Release CIRU Strix runtime v3.0.0 with qualified QSA and serving improvements

README.md

@@ -32,261 +32,115 @@ tags:
 
 [![Qwen3.8 Flash CIRU Strix IU4](assets/qwen38-flash-ciru-strix-iu4.jpg)](https://llm.ciru.ai/research)
 
-# Qwen3.8-Flash-CIRU-STRIX-IU4 · v2.0
+# Qwen3.8-Flash-CIRU-STRIX-IU4 · runtime v3.0.0
 
-**Mixed-precision Qwen3.8 for AMD Strix Halo, with ROCm 10 execution, NVMe-paged PLE, and depth-6 MTP.**
+**V3 brings faster long-context serving and the qualified QSA conversation-isolation fixes to the Strix-only runner. Model weights are unchanged.** It adds parallel attention-cell selection, indexed decode attention, cached derived history, guarded PLE lookups and wider prefill. The general profile retains maximum MTP depth 6 and batch/microbatch 1024.
 
-Version 2.0 brings together CIRU's retained ROCm 10 correctness and execution improvements. The measured strengths are **42.3 tok/s on the controlled non-thinking coding probe**, **382 tok/s cold prefill at 2K**, and **0.03045 mean forward KL to BF16** on the shared diagnostic panel. The target, MTP and PLE weights are unchanged.
+Use the [v3 source archive](runtime/v3.0.0/ciru-runtime-v3.0.0-source.tar.gz) or [matching GitHub tag](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v3.0.0). This text-only package requires the custom CIRU runtime, the target GGUF and all three `ple/` files. The `mtp/` head enables speculative decoding. Stock llama.cpp and Hugging Face hosted inference do not support this package.
 
-> [!IMPORTANT]
-> **Use the [v2.0 runtime archive](runtime/v2.0/ciru-runtime-v2.0-source.tar.gz) or matching [GitHub `v2.0` tag](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0).** This text-only package requires the custom CIRU runtime and all three `ple/` files. The `mtp/` draft enables speculation. Stock `llama.cpp` and Hugging Face hosted inference do not run this package correctly.
->
-> **The target GGUF, MTP draft and PLE weights are unchanged.** Existing users only need the new runtime and profile. v2.0 identifies the runtime/package release.
+## V3 serving comparison
 
-## Release results
+Measured on one Ryzen AI Max+ 395 / gfx1151 / 128 GB shared-memory NixOS machine, with one model workload at a time. Inputs are identical token IDs, cold prompt cache, 128 generated tokens, one slot and 262,144-token configured capacity. Nonthinking sampler: temperature 0.7, top-p 0.8, top-k 20, min-p 0, presence penalty 1.5, repeat penalty 1, frequency penalty 0 and seed 123. EOS is honored.
 
-| Result | CIRU v2.0 | Measurement |
-|---|---:|---|
-| Served non-thinking coding generation with MTP | **42.28–42.31 tok/s** | Two retained repetitions; 57-token prompt, 520 generated tokens, greedy coding profile |
-| Cold prefill at 2K | **382.03 tok/s** | 2,048 exact prompt tokens, MTP off |
-| Cold prefill at 128K | **232.95 tok/s** | 131,072 exact prompt tokens, MTP off |
-| Mean forward KL to BF16 | **0.03045** | 64 full-vocabulary distributions across four domains |
-| Observed-token PPL | **2.08401** | 60-token diagnostic slice; BF16 reference 2.01799 |
+| Input tokens | Profile | Prompt tok/s | Generation tok/s | Whole request (s) |
+| ---: | --- | ---: | ---: | ---: |
+| 4,096 | Previous CIRU | 392.00 | 22.52 | 16.34 |
+| 4,096 | CIRU v3 | 455.65 | 24.60 | 14.41 |
+| 4,096 | Halo | 381.49 | 35.30 | 14.69 |
+| 65,536 | Previous CIRU | 284.49 | 13.33 | 239.99 |
+| 65,536 | CIRU v3 | 369.81 | 24.22 | 182.57 |
+| 65,536 | Halo | 263.42 | 23.28 | 254.37 |
 
-The 42.3 tok/s result uses an explicitly closed thinking block and our retained fixed-six configuration: ROCm10, a 32,768-row draft shortlist, F16 target KV, Q8_0 draft KV, 16K context and the performance CPU governor. Both runs produced the same 520 output tokens. **It is a controlled coding-probe rate, not a HumanEval average or a claim that every workload runs at that speed.** [Recorded coding results](benchmarks/v2.0/controlled-coding-42tps.json).
+MTP 2 is a useful optional setting for the tested lower-acceptance long requests, but it was not promoted as the general default. Across 20 short HumanEval requests, v3 MTP 6 measured 53.24 generation tok/s versus 39.63 with MTP 2. Both passed 20/20 base and extended tests. Keeping MTP 6 avoids that short-coding regression; target verification retains the full vocabulary at either depth.
 
-The comparison preserves each package's own recorded runtime settings. On the shared MTP-off sweep, CIRU has higher prefill from 2K through 128K: at 128K it is **26.4% faster than Agention/Laurent and 22.7% faster than Unsloth**. On the small BF16 panel, mean forward KL is **72.9% lower than Agention/Laurent and 82.6% lower than Unsloth**. These are fidelity and prefill measurements, not overall task-quality or generation-speed percentages.
+| Input tokens | Optional v3 MTP 2 prompt tok/s | Generation tok/s | Whole request (s) |
+| ---: | ---: | ---: | ---: |
+| 4,096 | 453.09 | 29.39 | 13.62 |
+| 65,536 | 373.08 | 24.88 | 180.87 |
 
-## Three-way comparison
+The previous CIRU arm is the locally qualified v2.0.1 runner under its original MTP 6, b2048/u512 profile. The v3 arm uses the same weights and MTP 6, with b1024/u1024. Halo is the unmodified current fork at commit `5f851647fe5ed795dfd6c0a3fba543114879e874`, using its recommended Vulkan backend, Unsloth UD-Q4_K_XL target and published EasiiX Strix Q8 MTP head. Native KV, batch, thread, fitting and cache defaults are retained.
 
-The three-way context sweep ran on **Sozo: Ryzen AI MAX+ 395 / Radeon 8060S, gfx1151, 128 GiB shared memory, NixOS**, with one model workload at a time. Ciru handled the new CIRU BF16 captures and the clean Ubuntu build/GPU smoke. The comparisons preserve each arm's recorded execution settings; they are package comparisons, not a controlled kernel-only experiment.
+Halo maximum depths 2, 3, 4, 6 and native adaptive 6 were screened. Depth 3 won its short-context screen at **35.37 tok/s**, versus **31.4** at depth 2, **30.04** at depth 4, **25.91** at depth 6 and **29.31** with adaptive 6. The final comparison above uses depth 3. Halo source and weights were not modified.
 
-| Arm | Target and runner | MTP implementation available in the package |
-|---|---|---|
-| CIRU v2.0 | Released IU4 weights; locked RC2 ROCm 10 runtime | Fixed maximum 6, p-min 0, 32,768-row draft shortlist; target F16 KV, draft Q8_0 KV |
-| Agention / Laurent | [FP4 FAST model](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix-GGUF/tree/ad4c5717254a630ee0c5a8db5208eb1f8476e56c); [Laurent's Vulkan fork](https://github.com/LaurentZuijdwijk/llama.cpp/commit/5e085d123eead2e89b5c19f824fccb05727da6a2) | Publisher adaptive 2–4 with its FP4 draft; target Q8_0 KV |
-| Unsloth / recommended | [IQ4_XS model](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF/tree/38bb39ee97821de2c9009abb7e93950eec396e66); [Daniel Han Chen's MTP branch](https://github.com/danielhanchen/llama.cpp/commit/d1a92352cbd417fd840b4e765c0b82f5fe3d1d89), native Vulkan build | Unsloth [MTP README option 2](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF/blob/38bb39ee97821de2c9009abb7e93950eec396e66/MTP/README.md), maximum 2 and shared Q8_0 draft; publisher defaults |
+CIRU and Halo have different quantizations and execution profiles: this is a serving-package comparison. Generation throughput, prompt processing and whole-request latency are separate metrics. At 4K, the general MTP 6 profile is close to Halo in total time; the optional MTP 2 setting provides the clearer latency benefit on this fixture. Long-context prompt processing shows the larger gain. The tables retain Halo's generation advantage where present; a CIRU request-time win is not a claim of winning every metric. Three clean v3 loads, two selected Halo loads and one previous-CIRU load are a bounded experiment, not a confidence interval or general ranking.
 
-Only these three combinations are included. Exact model/shard hashes and runner commits are in [competitor artifacts](benchmarks/v2.0/competitor-artifacts.json); complete recorded settings are in [runtime recipes](benchmarks/v2.0/runtime-recipes.json).
+The 2.79 GB Unsloth shared Q8 head intentionally omits tensors a supporting loader borrows from the main model. The pinned Halo loader fails for missing `token_embd.weight`; Unsloth's self-contained Q8 head also fails for missing `output_hc_norm.weight`. Both attempts are recorded. The compatible [EasiiX Strix Q8 head](https://huggingface.co/EasiiX/Qwen3.8-Flash-Next-MTP-Strix-Halo-GGUF/tree/6f7900648b1c6b14f067a182c640e47971e9ab35) is used as published.
 
-### BF16 numerical fidelity
+[Full report, first-piece latency and memory](benchmarks/v3.0.0/COMPARISON.md) · [Structured results](benchmarks/v3.0.0/comparison.json) · [Raw evidence archive](benchmarks/v3.0.0/strix-v3.0.0-evidence.tar.gz)
 
-![BF16 fidelity diagnostic](assets/v2-bf16-fidelity.png)
+## Quality and capacity checks
 
-| Configuration | Mean forward KL ↓ | p95 KL ↓ | BF16 top-token agreement | Observed-token PPL ↓ |
-|---|---:|---:|---:|---:|
-| BF16 reference | 0 | 0 | 64/64 | 2.01799 |
-| CIRU ROCm10 v2.0 | **0.03045** | 0.15453 | 61/64 | 2.08401 |
-| Agention / Laurent | **0.11223** | 0.29765 | 60/64 | 2.38380 |
-| Unsloth / recommended | **0.17457** | 0.55722 | 61/64 | 2.42931 |
+| Profile | HumanEval base | EvalPlus extended tests | Recall at about 8K and 64K |
+| --- | ---: | ---: | --- |
+| Previous CIRU | 20/20 | 20/20 | Both keys and exact cached replay |
+| CIRU v3 | 20/20 | 20/20 | Both keys and exact cached replay |
+| Halo | 20/20 | 20/20 | Both keys and exact cached replay |
 
-This fixed diagnostic contains **64 full-vocabulary distributions and 60 observed next-token losses** across four short domain prefixes. MTP is off. CIRU uses F16 KV and flash-attention auto; the retained competitor captures use Q8_0 KV and flash attention on. CIRU's two independent model loads produced byte-identical logits.
+These are canonical HumanEval tasks 0–19, EvalPlus v0.1.10, one first sample per task, no retries and a 4096-token cap; truncations fail. Generated code runs inside a filesystem/network sandbox. This small nonthinking coding and recall panel is a regression check. It does not establish broad model equality, thinking-mode quality, tool reliability or leaderboard standing.
 
-The result measures these implementations' fidelity on a small shared panel. It does not establish general task-quality superiority or full-corpus perplexity. The old CIRU/Q5 diagnostic used an earlier runtime and is superseded for the current three-way comparison.
+V3 also completed **261,888 input tokens plus 128 generated tokens** at **257.44 prompt tok/s and 18.00 generation tok/s**, with a **1024.44 s** whole request. This is a CIRU-only serving-capacity check, not a filled-256K Halo comparison or full-context accuracy result.
 
-[Full statistics and hashes](benchmarks/v2.0/bf16-fidelity.json) · [CSV](benchmarks/v2.0/bf16-fidelity.csv)
+The final inference source passes 69 QSA mapping/state/guard cases with flags on and off, 33 actual ROCm operator reference cases and the existing 30 batch allocator tests. A separate short four-prefix diagnostic matches 15,892,480 F32 logits byte-for-byte against the previous runner. Radix selection can change threshold-tie membership and selected-list order; general bitwise equivalence is not claimed.
 
-### MTP-off context sweeps
+## Download, build and run
 
-![MTP-off context sweeps](assets/v2-context-sweep.png)
+The unchanged model files total **135,962,881,135 bytes (126.625 GiB)**, excluding runtime and reports. The target is 73.945 GiB, the PLE payload 48.828 GiB and the Q8 MTP head 3.852 GiB. Allow at least 160 GiB for model storage/verification plus SDK and build space. Use fast NVMe and a Strix Halo machine with 128 GiB unified memory.
 
-Each cell is **prefill / generation tok/s**. All seven points use the same exact-token source fixture, a cold prompt, and 128 generated tokens.
-
-| Prompt tokens | CIRU ROCm10 v2.0 | Agention / Laurent | Unsloth / recommended |
-|---:|---:|---:|---:|
-| 512 | 306.53 / 22.55 | 357.63 / 25.93 | 235.03 / 24.54 |
-| 2,048 | 382.03 / 20.94 | 363.83 / 25.58 | 270.12 / 23.98 |
-| 8,192 | 370.41 / 19.19 | 325.59 / 24.95 | 274.78 / 22.69 |
-| 16,384 | 352.13 / 17.47 | 302.72 / 24.36 | 266.87 / 21.28 |
-| 32,768 | 321.33 / 14.32 | 273.95 / 22.77 | 254.68 / 18.79 |
-| 65,536 | 282.99 / 10.17 | 229.68 / 20.25 | 228.78 / 13.37 |
-| 131,072 | 232.95 / 6.75 | 184.33 / 17.86 | 189.79 / 9.98 |
-
-Server context is 262,144, sampling is greedy with seed 1234, EOS is ignored for the 128-token measurement, and a 512+32 warmup is excluded. Each row verifies exact prompt/output counts and zero drafted/accepted tokens.
-
-CIRU leads prefill at 2K and above. **Both competitor arms have higher MTP-off generation rates across this sweep.** These are target-only context measurements, separate from the native MTP panel. Unsloth's target-only sweep uses Q8_0 KV, 16 threads, batch 2048/microbatch 512 and explicit CPU PLE placement with lazy mode off; its native MTP panel uses publisher defaults. CIRU uses F16 KV, 8 threads, and its 4 GiB PLE cache. See the recorded recipes for all differences.
-
-[Full sweep CSV](benchmarks/v2.0/mtp-off-context-sweep.csv) includes TTFP and memory. [Structured results](benchmarks/v2.0/comparison.json) preserve idle/peak/delta RAM, GTT and VRAM counters. RAM is whole-system usage; these overlapping shared-memory counters must not be summed.
-
-
-
-## What changed in v2.0
-
-The locked candidate includes the retained Sozo quality-lane and Ciru speed-lane fixes:
-
-- Correct attention descriptor strides, GPU admission for long QSA top-k, and restored-KV fast-path admission.
-- Seven-column Q5_K weight reuse, the GPU MTP top-10 selector and a 32,768-row draft-output shortlist.
-- Decode-entry and internal-microbatch synchronization repairs.
-- Fresh HIP graph recapture, plus a launcher working-directory fix.
-- The complete ROCm SDK installation helpers from v1.1.1, with the updated depth-6 production profile.
-
-The previous allocator-lifetime and cached-prefix/MTP state fixes remain included. The adaptive runner and broad expert-reuse/fusion experiments are not enabled in this release. No weights were requantized or retrained.
-
-The core source matches **`qwen38-ciru-rocm10-20260905-rc2`**. Portable build helpers are based on public v1.1.1 (`764ee491`); the retained core patch is based on v1.1 (`baba5e06`). [Provenance and runtime hashes](benchmarks/v2.0/ciru-provenance.json) record the exact inclusion evidence. The additional long-QSA/restored-KV admission code does not by itself establish a measured production cache-recovery speedup.
-
-## Download and build
-
-The model artifacts total **135,962,881,135 bytes (126.625 GiB)**, excluding runtime source and reports. Target storage is 73.945 GiB, the MTP draft is 3.852 GiB, and the mandatory PLE payload is 48.828 GiB. The target and all three `ple/` files are required; the MTP draft is optional only when speculation is disabled.
-
-On Ubuntu/Debian, install `python3-venv`, then download the package:
+On Ubuntu/Debian, install Git and a Python virtual environment, then download the model and build the source:
 
 ```bash
-sudo apt update
-sudo apt install -y python3-venv
+sudo apt-get update
+sudo apt-get install -y git python3-venv
 python3 -m venv .venv-hf
 .venv-hf/bin/python -m pip install -U huggingface_hub
 . .venv-hf/bin/activate
 hf download jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 \
-  --revision v2.0 --local-dir ./Qwen3.8-Flash-CIRU-STRIX-IU4
-cd Qwen3.8-Flash-CIRU-STRIX-IU4
-sha256sum -c checksums.sha256
-sha256sum -c v2.0-checksums.sha256
-tar -xzf runtime/v2.0/ciru-runtime-v2.0-source.tar.gz
-cd ciru-runtime-v2.0
-./scripts/ciru/setup-linux-amd.sh --install-host-deps
-```
-
-The helper installs host build tools and a private, complete **ROCm 10.0.0** SDK with gfx1151 device libraries. It builds into `build-gfx1151-sdk/`. Keep `.venv-rocm/` for runtime libraries. The host must already provide a compatible AMD GPU driver and access to `/dev/kfd` and its render node.
-
-**Existing users:** keep your model directory. Download only the runtime archive, updated profile and release checksums into a separate directory, then rebuild:
-
-```bash
-hf download jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 \
-  --revision v2.0 --local-dir ./ciru-v2-update \
-  --include "runtime/v2.0/*" "profiles/*" "v2.0-checksums.sha256"
-cd ciru-v2-update
-sha256sum -c v2.0-checksums.sha256 --ignore-missing
-tar -xzf runtime/v2.0/ciru-runtime-v2.0-source.tar.gz
-cd ciru-runtime-v2.0
+  --revision v3.0.0 --local-dir ./model
+(cd model && sha256sum -c checksums.sha256 && sha256sum -c v3.0.0-checksums.sha256)
+git clone --branch v3.0.0 --single-branch \
+  https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4.git ciru-runtime-v3.0.0
+cd ciru-runtime-v3.0.0
 ./scripts/ciru/setup-linux-amd.sh --install-host-deps
-```
-
-The source archive includes build/run instructions, provenance, licenses and third-party notices. The matching [GitHub `v2.0` tag](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0) now provides the same source in Git: all 3,537 file contents and Git modes match the published archive. [Release downloads](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/releases/tag/v2.0) · [Diff from v1.1.1](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/compare/v1.1.1...v2.0) · [Recorded source identity](benchmarks/v2.0/git-source.json).
-
-As an alternative to extracting the archive, with Git installed:
-
-```bash
-git clone --branch v2.0 --single-branch https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4.git ciru-runtime-v2.0
-cd ciru-runtime-v2.0
-./scripts/ciru/setup-linux-amd.sh --install-host-deps
-```
-
-The original archive and Hugging Face `v2.0` revision remain unchanged. Later documentation clarifications and an early multi-slot MTP launcher check appear on GitHub `main`; the released inference code is unchanged. Use the [current running instructions](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md) for profile verification and draft-depth guidance.
-
-Benchmarked on NixOS with ROCm10. The runtime also passed an Ubuntu 24.04 container build and GPU smoke test.
-
-## Run the production profile
-
-From the cloned or extracted runtime directory:
-
-```bash
-BUILD_DIR="$PWD/build-gfx1151-sdk" \
-  MODEL_DIR=/absolute/path/to/Qwen3.8-Flash-CIRU-STRIX-IU4 \
+BUILD_DIR="$PWD/build-gfx1151-sdk" MODEL_DIR="$(realpath ../model)" \
   ./scripts/ciru/run-server.sh
 ```
 
-The profile uses one slot, a 262,144-token context, batch 2048/microbatch 512, 8 target threads, F16 target KV, Q8_0 draft KV, flash attention on, maximum MTP depth 6, p-min 0 and p-split 0.10. It enables a separate 4 GiB PLE cache with 16 workers, an 8 GiB prompt cache, idle-slot caching and context checkpoints. The server binds to `127.0.0.1:8080`.
+The helper installs a private complete ROCm 10.0.0 SDK with gfx1151 libraries. Keep the SDK directory used by the build; the host needs a compatible AMD driver and access to `/dev/kfd` and its render node. New v3 GPU qualification is NixOS/ROCm 10/gfx1151. The prior clean Ubuntu qualification belongs to v2.0.1, so the Ubuntu instructions are retained build guidance, not a new v3 Ubuntu test result. See [platform/build instructions](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v3.0.0/docs/BUILD_LINUX.md).
 
-### Confirm the MTP profile
+Existing users can keep their model directory and clone/build only the new runtime. The source archive is equivalent to the Git tag including executable modes and symlinks. The optional [tested NixOS binary payload](runtime/v3.0.0/ciru-runtime-v3.0.0-nixos-gfx1151.tar.gz) requires the recorded Nix store and ROCm SDK paths; use the source build for another installation.
 
-The launcher exports `profiles/strix-halo-production.env`, including `CIRU_MTP_SHORTLIST=32768` and `CIRU_MTP_TOPK10=1`. Compiling v2.0 and setting `--spec-draft-n-max 6` alone does not enable those optimizations. With MTP enabled, confirm this startup line:
+The launcher enables 262,144 context capacity, F16 target KV, Q8 draft KV, the 32,768-row draft shortlist, maximum MTP depth 6, b1024/u1024, eight CPU threads and prefix caching. MTP uses `--parallel 1`; multi-slot MTP is rejected before model load. For target-only parallel serving, set `ENABLE_MTP=0` and follow the [parallel instructions](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v3.0.0/docs/RUNNING.md#parallel-requests-and-unified-kv-cache).
 
-```text
-CIRU MTP shortlist enabled: 32768 / 248320 vocabulary rows; full target verification retained
-```
-
-For direct `llama-server` launches, use the [complete exported environment and flags](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md#2-launch-the-audited-public-profile).
+Confirm `CIRU MTP shortlist enabled: 32768 / 248320 vocabulary rows; full target verification retained` in the startup log. The shortlist limits draft projection; target verification retains the full vocabulary. `MTP_DEPTH=2` selects the tested option for low-acceptance long requests; the screen does not establish the optimum for every prompt. New v3 optimization switches accept literal `0`. The optional draft attention window remains off and is unqualified when enabled.
 
-### Choose a draft depth
-
-**Depth 6 is the released default; the fastest depth depends on the workload.** When later draft tokens are frequently rejected, depth 3 may be faster, including for some short interactive completions. Restart with this command to change only the maximum depth:
-
-```bash
-MTP_DEPTH=3 BUILD_DIR="$PWD/build-gfx1151-sdk" \
-  MODEL_DIR=/absolute/path/to/Qwen3.8-Flash-CIRU-STRIX-IU4 \
-  ./scripts/ciru/run-server.sh
-```
-
-The controlled coding result above does not establish that six is optimal for short chat, long contexts or batched serving. The context sweeps have MTP off and do not compare draft depths. No adaptive depth selection is enabled. Single sampled completions are useful directional feedback, but need matched settings and repeated measurements to support a general speed claim. See [comparison details](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md#confirm-the-mtp-profile-and-choose-a-draft-depth).
-
-### Send a chat request
+Thinking mode remains the model default: temperature 1.0, top-p 0.95, top-k 20, min-p 0. For the nonthinking mode evaluated here:
 
 ```bash
 curl http://127.0.0.1:8080/v1/chat/completions \
-  -H 'Content-Type: application/json' \
+  -H "Content-Type: application/json" \
   -d '{
     "model": "Qwen3.8-Flash-CIRU-STRIX-IU4",
-    "messages": [{"role": "user", "content": "Hello"}],
-    "temperature": 1.0,
-    "top_p": 0.95,
-    "top_k": 20,
-    "min_p": 0,
-    "cache_prompt": true
+    "messages": [{"role": "user", "content": "Write a Python CSV validator with tests."}],
+    "chat_template_kwargs": {"enable_thinking": false},
+    "temperature": 0.7, "top_p": 0.8, "top_k": 20,
+    "min_p": 0, "presence_penalty": 1.5, "cache_prompt": true
   }'
 ```
 
-For target-only serving, set `ENABLE_MTP=0` before launching. This omits the draft flags. Use the launcher option rather than appending `--spec-type none` to an already speculative command.
-
-Thinking defaults are temperature 1.0, top-p 0.95, top-k 20, min-p 0. For non-thinking mode, send `"chat_template_kwargs":{"enable_thinking":false}` with temperature 0.7, top-p 0.8 and presence penalty 1.5. Normal chat can retain prefix caching; the cold benchmark settings are not the production preset.
-
-`min_p=0` explicitly follows [Qwen's model-specific recommendation](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/f5d08274bafd880402bd16f5e3e6c514136ec06c/README.md#best-practices); v1.1 could inherit llama.cpp's `0.05`. This disables an additional probability filter and can change sampled output. We have not established that either value gives better answer quality in a controlled A/B test. Set `MIN_P=0.05` with the launcher to select the previous inherited value; request-level settings can override the server default.
-
-### Parallel agents and unified KV
-
-**The released MTP shortlist supports one slot only.** For two slots, disable MTP explicitly; otherwise the original v2.0 runtime asserts during initialization. The current GitHub `main` launcher now catches this before loading the model. The original `v2.0` tag and archive remain unchanged.
-
-From either the original v2.0 runtime directory or the current repository:
-
-```bash
-ENABLE_MTP=0 PARALLEL_SLOTS=2 CONTEXT_SIZE=524288 \
-  BUILD_DIR="$PWD/build-gfx1151-sdk" MODEL_DIR=/absolute/path/to/model \
-  ./scripts/ciru/run-server.sh --no-kv-unified
-```
-
-With separate KV caches, `524288` is the total context allocation: **262144 per slot**, not 512K per agent. Inspect `/slots` for the actual limits.
-
-The released CIRU/ROCm 10 runtime passed 16 target-only requests in our 2026-09-06 two-slot smoke: eight with separate KV and eight with unified KV, including overlapping short/long and streamed requests. This does not validate filled 512K contexts or every agent workload. The community subsequently identified [#27994](https://github.com/ggml-org/llama.cpp/issues/27994), the Qwen3.8/QSA unified-cache sequence-isolation bug fixed upstream by [#27941](https://github.com/ggml-org/llama.cpp/pull/27941). **The relevant fix is missing from our v2.0 source. Keep `--no-kv-unified` explicit for multi-slot serving pending integration and validation.** The short marker tests do not validate conversation recall and cannot rule out that bug. The launcher guard addresses the separate one-slot MTP assertion. The HIP host-buffer candidate discussed in the validation report is a different issue and remains unshipped.
+These sampling defaults follow the [Qwen model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/f5d08274bafd880402bd16f5e3e6c514136ec06c/README.md#best-practices). Ordinary chat can retain prefix caching; the benchmark's cold requests, fixed seed and output cap are measurement controls.
 
-[Two-slot instructions and upstream issue links](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md#parallel-requests-and-unified-kv-cache) · [Validation report](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/PARALLEL_VALIDATION.md)
+## What ships in v3
 
-## Historical task-quality results on the unchanged weights
+V3 combines the previously qualified QSA sequence isolation/indexer-copy fixes with the portable retained hybrid work: wider aligned QSA prefill, J32 expert tiling for supported prefill shapes, radix selection, indexed F16 attention, direct guarded PLE lookup, derived-history caching keyed by sequence and packed tiny F32 gathers. The F32 derived cache adds about 192 MiB at 256K. Hybrid external-GPU ownership, peer transfers, expert splitting, scheduler overlap and experimental compact masks are excluded.
 
-These earlier H96/depth-1 results were **not rerun as full suites on v2.0**. The current release evidence comprises the bounded BF16 panel, controlled coding probe, context sweeps, and scoped runtime correctness/build checks described above.
+The original READY package, source, evidence and checksum trees are included in the [prior-package archive](benchmarks/v3.0.0/qsa-v2.0.1-prior-package.tar.gz). Its v2.0.1 fixes ship as part of v3; no separate public v2.0.1 tag is claimed. Existing public v2.0 tags and weight identities remain unchanged. [Historical v2.0 results](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0/README.md) remain available; their 42.3 tok/s short greedy probe and broader older quality suites are different workloads and were not rerun as full suites for v3.
 
-| Benchmark | Earlier result | Scope |
-|---|---:|---|
-| HumanEval | 160/164 (97.56%) | Full local-custom chat run, first sample |
-| HumanEval+ | 155/164 (94.51%) | Full EvalPlus 0.1.10 run |
-| ARC-Challenge | 1,143/1,172 (97.53%) | Full EvalScope dataset |
-| ToolEval Standard | 115/138 (83.33%) | 69 local-custom cases |
-| ToolEval Hard | 23/30 (76.67%) | 15 local-custom cases |
+The IU4 model name is retained. The target has 1,223 tensors: 144 Q4_1 routed-expert tensors, 328 Q5_K, 290 Q8_0, 48 Q5_1, 25 BF16 and 388 F32. The standard launcher uses ordinary GGUF types; its Q4_1 matrix path expands packed values into byte lanes for IU8 WMMA. It does not activate the separate native IU4/E3 bank path. The PLE payload preserves exact FP8 E4M3 weights with one BF16 scale.
 
-Historical methodology and additional sampled results remain available in the [previous model card](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/9327d8c4b796d983d97870f0908e8667c0cbab2d/README.md). Local-custom or sampled scores are not canonical leaderboard submissions.
-
-## Storage and hardware
-
-The target contains 1,223 tensors: 144 routed-expert Q4_1 tensors, 328 Q5_K, 290 Q8_0, 48 Q5_1, 25 BF16 and 388 F32. **IU4 describes the optimized unsigned four-bit execution path**, not a uniform model quantization or a new GGUF storage format. The PLE payload preserves exact FP8 E4M3 weights with one BF16 scale.
-
-The intended machine is a Ryzen AI MAX+ 395 / Radeon 8060S with 128 GiB unified memory and fast NVMe. Allow at least 160 GiB for model files and verification, plus space for the source, SDK and build. Linux provides the measured P16/O_DIRECT pager path. The configured context is 262,144; cold sweep coverage stops at 131,072.
-
-[File tree](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0) · [Weight checksums](checksums.sha256) · [Runtime/report checksums](v2.0-checksums.sha256) · [Measurement methodology](benchmarks/v2.0/METHODOLOGY.md)
+[Model file tree](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v3.0.0) · [Weight checksums](checksums.sha256) · [V3 runtime/report checksums](v3.0.0-checksums.sha256) · [Source identity](benchmarks/v3.0.0/git-source.json) · [GitHub release](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/releases/tag/v3.0.0)
 
 ## Lineage, license and credit
 
-- Text lineage: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/tree/f5d08274bafd880402bd16f5e3e6c514136ec06c).
-- PLE lineage: [Qwen/Qwen3.8-Flash-Next-FP8](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8/tree/bcd9f01ddc9cff2316eb84281bebcd5b058bddce).
-- Runtime base: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/commit/f5e85d43a048f3d5adefb4c5e29867d8077fba62).
+Text lineage is [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/tree/f5d08274bafd880402bd16f5e3e6c514136ec06c); PLE lineage is [Qwen3.8-Flash-Next-FP8](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8/tree/bcd9f01ddc9cff2316eb84281bebcd5b058bddce). Runtime lineage starts from [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/commit/f5e85d43a048f3d5adefb4c5e29867d8077fba62). Model artifacts use the included Qwen Community License 1.0; runtime code retains MIT and component notices.
 
-Model artifacts use the included **Qwen Community License 1.0**; runtime code retains its MIT license and component notices. Credits include Qwen, ggml-org, Ryan Monsurate's MTP integration, AMD's ROCm ecosystem, and contributors recorded in the source notices. Thanks also to Laurent Zuijdwijk, Agention AI, Unsloth and Daniel Han Chen for the publicly available comparison models and runtimes, and OpenAI's HumanEval authors for the benchmark tasks.
-
-CIRU is an independent community research project. AMD and Qwen marks do not imply sponsorship or endorsement.
-
-```bibtex
-@software{ciru_qwen38_flash_strix_iu4_2026,
-  title   = {Qwen3.8-Flash-CIRU-STRIX-IU4},
-  author  = {{CIRU}},
-  year    = {2026},
-  version = {2.0},
-  url     = {https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4}
-}
-```
+Credits include Qwen, ggml-org, Ryan Monsurate's MTP integration, AMD's ROCm ecosystem and the contributors in the source notices. Daniel Han's authorship of the imported QSA repair is preserved. Thanks to halo-box, EasiiX, Unsloth, Daniel Han Chen, Laurent Zuijdwijk and Agention AI for public models/runtimes, and the HumanEval and EvalPlus authors for evaluation tools. CIRU is an independent community research project; AMD and Qwen marks do not imply sponsorship.

benchmarks/v2.0.1/FINAL-QUALIFICATION.json

@@ -0,0 +1,94 @@
+{
+  "status": "QUALIFIED_FOR_SCOPED_RELEASE",
+  "release": "v2.0.1",
+  "decision_utc": "2026-09-07T23:06:35.606923+00:00",
+  "source_code_commit": "bf0bf4b795e112bb1b4af0101640b8fdb250365f",
+  "scope": "Focused QSA conversation isolation/indexer state backport plus guarded canonical input construction. Two-slot tests use MTP disabled; production MTP remains one slot.",
+  "correctness": {
+    "qsa_mapping_state_cases": 66,
+    "canonical_input_tensor_cases": 288,
+    "clean_batch_tests": 30,
+    "clean_batch_assertions": 198,
+    "served_recall_passed": 16,
+    "served_recall_total": 16,
+    "active_initial_recall_prompt_tokens": 13062,
+    "separate_and_unified_kv": true,
+    "full_model_controls": "Cache copies/state round trips, equal-layout foreign-content and simultaneous-sequence controls, normalization boundary checks; see original structured summaries."
+  },
+  "performance": {
+    "gate": "Less than 2% PP/TG loss in the final bounded pair, exact target tokens and MTP counts; inspect hardware telemetry for material mismatches.",
+    "ciru_real_64k": {
+      "prompt_tokens": 65536,
+      "generated_tokens": 128,
+      "loads": 2,
+      "order": [
+        "candidate",
+        "baseline"
+      ],
+      "changes_percent": {
+        "prompt_per_second": 2.1206494168475754,
+        "predicted_per_second": -0.44880462382312913
+      },
+      "tokens_exact": true,
+      "draft_counts_exact": true,
+      "telemetry_review": "No material clock, power, CPU or I/O anomaly; baseline was modestly warmer. No causal attribution of the older slowdown."
+    },
+    "sozo_real_8k": {
+      "prompt_tokens": 8192,
+      "generated_tokens": 128,
+      "loads": 2,
+      "order": [
+        "candidate",
+        "baseline"
+      ],
+      "prefill_change_percent": 4.699006735082145,
+      "generation_change_percent": 1.2031817171471504,
+      "tokens_exact": true,
+      "draft_counts_exact": true
+    },
+    "short_coding": {
+      "generated_tokens": 520,
+      "tokens_exact": true,
+      "pooled_tg_change_percent": 0.10079194713958906
+    },
+    "historical_sozo_64k": "Earlier pair showed PP -2.61% / TG -6.86%; original evidence retained, cause unresolved. Final Ciru pair does not turn it into a passing Sozo result.",
+    "profiler": "Attachment failed before any diagnostic request. No GPU kernel trace captured; final decision uses clean served timing, native counts, telemetry and prior source-level controls."
+  },
+  "clean_build": {
+    "status": "PASS",
+    "userspace": "Ubuntu 24.04",
+    "sdk": "ROCm 10.0.0",
+    "build_host": "Dunamis CPU only",
+    "cpu_isa": "Explicit AVX2 portable configuration, GGML_NATIVE=OFF, AVX512/VNNI disabled",
+    "gpu_smoke_host": "Ciru gfx1151 using NixOS host driver",
+    "token_sha256": "5115118aef25826d9d904cd30f47a28fc9aeac842457b17ff906deba653fb731",
+    "matched_tokens": 520,
+    "draft_n": 674,
+    "draft_n_accepted": 400,
+    "gpu_smoke_receipt": "ubuntu24-gpu-smoke/REVIEWED-RESULT.json",
+    "original_verifier_error": "Literal gfx1151 omitted by low-verbosity log; corrected using device enumeration/KFD identity and actual RDNA3.5 execution log, without inference retry."
+  },
+  "unchanged": [
+    "Target and draft weights",
+    "PLE sidecars",
+    "HIP sources",
+    "Sampling defaults",
+    "Production MTP depth and shortlist"
+  ],
+  "excluded": [
+    "HIP host-buffer #25863",
+    "Kairic",
+    "HC-mix",
+    "Multi-slot MTP support",
+    "Native-IU4 E3 experiments"
+  ],
+  "limits": [
+    "No filled 512K context or every-agent-workload claim",
+    "No new full task-quality suite",
+    "Physical cache layouts can change floating-point results",
+    "Ubuntu container userspace smoke is not native Ubuntu host certification",
+    "Single matched performance pair per new context, not statistical confidence bounds"
+  ],
+  "cleanup": "Ciru GPU released; no owned llama-server or running test containers; service remains inactive and performance governors unchanged. Dunamis build containers stopped.",
+  "publication_state_at_decision": "Local release preparation; no upload or deployment performed"
+}

benchmarks/v2.0.1/QUALIFICATION.md

@@ -0,0 +1,100 @@
+# QSA sequence-isolation fix: v2.0.1 qualification
+
+Qualification completed on 2026-09-07. Correctness and scoped performance checks passed. The clean Ubuntu build and GPU smoke passed, including exact reference output and MTP counts.
+
+v2.0.1 fixes the QSA conversation-isolation defect reported in [upstream #27994](https://github.com/ggml-org/llama.cpp/issues/27994). It includes the relevant sequence grouping, indexer copying, saved-position metadata and normalization-shape changes from Daniel Han's [#27941](https://github.com/ggml-org/llama.cpp/pull/27941), commit `36b10154383b60eb15baac2c7a40d2a5f784faa7`. Unrelated upstream metadata validation and tensor-parallel policy changes are outside this backport.
+
+QSA now pools cache cells by both sequence set and position bucket, so a second conversation with overlapping token positions cannot replace the first conversation's indexer blocks. Pending stream copies also update the raw indexer keys. A guarded path for one contiguous sequence emits the same QSA inputs with less host work. Shared prefixes, fragmented caches, multiple sequences and repeated spatial positions use the general sequence-aware path.
+
+Weights, HIP kernel sources, sampling defaults, MTP depth and the shortlist policy are unchanged. The retained NixOS HIP binary is unchanged; the clean Ubuntu build has its own binary identities. Multi-slot MTP is still unsupported: use `ENABLE_MTP=0 PARALLEL_SLOTS=2` for two slots.
+
+## Correctness
+
+| Check | Result and scope |
+| --- | --- |
+| Clean CPU source build | Passed; 66 mapping/state cases and the existing batch-allocation test passed |
+| Original runtime mapping reproducer | v2.0 failed 16/16 overlapping-sequence cases; the backport passed all |
+| Guarded path versus general reference | All four QSA input tensors matched by SHA-256 across 288 cases, including offsets, tails, two streams and both bias formats |
+| Full-model cache copy and restore | Split-KV copying fixed; all 16 reference logit vectors exact; state roundtrips byte-exact in split and unified modes |
+| Equal-layout foreign-content control | v2.0 failed; v2.0.1 returned byte-identical full-vocabulary logits for all 16 steps |
+| Simultaneous two-sequence decode batches | Changing the second conversation's complete prompt and suffix left the first conversation's 16 full-vocabulary vectors byte-identical in split and unified KV; v2.0 failed unified KV |
+| Actual served recall | 16/16 requests passed: eight split, eight unified; 13062-token initial prompts with the private code near the beginning, simultaneous requests, followups, slot reuse, prefix caching, streamed and nonstreamed responses |
+| GPU normalization boundary | Passed 65535, 65536 and 131072 pooled rows on ROCm0; all finite, maximum absolute error 3.61e-7 against a double reference |
+
+Full-model concurrency tests used 524288 configured total context, two slots, MTP disabled, F16 target KV and a 4096 MiB PLE cache. Both served slots reported a 262144-token limit. These are bounded isolation and recall tests, not a filled-512K run or a complete agent-workload qualification.
+
+Different physical cache layouts can still produce different floating-point results. The original solo-versus-shared gate failed for both builds and is retained in the evidence. A tensor trace located the first difference in the existing flash-attention calculation when the physical cell window grew from 6400 to 10496; inputs and logical selected cells matched at that point. The separately specified equal-layout control isolates foreign-conversation content from that layout effect. We do not claim bitwise invariance across different physical layouts.
+
+## Performance
+
+Serving comparisons of the final guarded candidate used the retained v2.0 RC2 baseline and the same QSA candidate library, ROCm 10.0.0, 262144 configured context, one slot, F16 target KV, Q8_0 draft KV, maximum draft depth 6, shortlist 32768 and performance CPU governors. Target, draft and PLE identities were held fixed. Sozo and Ciru are separate Strix Halo gfx1151 hosts with Ryzen AI MAX+ 395, Radeon 8060S and 128 GiB shared memory, running NixOS. Comparisons are paired within each host; results from different machines are not pooled.
+
+The short served MTP check on Sozo used four fresh loads (A1/C1/C2/A2), a frozen 57-token coding request and 520 greedy output tokens. Every output token and draft/acceptance count matched. Geometric pooled generation throughput changed by **+0.10%**, passing the prespecified less-than-2% loss and 3-percentage-point mirror-gap screen.
+
+Two subsequent confirmations used exactly one fresh candidate load and one fresh baseline load each, with the same 512-token prompt / 16-output warmup excluded from timing. Both used cold exact-token prompts and 128 greedy output tokens:
+
+| Host / date | Actual prompt tokens | Baseline prefill / generation tok/s | Candidate prefill / generation tok/s | Prefill change | Generation change |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Sozo / 2026-09-06 | 8192 | 370.5067 / 38.5802 | 387.9168 / 39.0444 | +4.6990% | +1.2032% |
+| Ciru / 2026-09-07 | 65536 | 284.1161 / 17.1228 | 290.1413 / 17.0459 | +2.1206% | -0.4488% |
+
+Both confirmations passed the exact-output/work-count and less-than-2% throughput-loss gates. The Ciru pair returned the same 128 tokens, with 225 drafted / 89 accepted tokens and 39 target graph reuses in each arm. The official recorder retained raw SSE, request/settings, token IDs, memory samples, JSONL/SQLite records and verified hash ledgers. These are bounded regression checks, not general speedup claims.
+
+Ciru telemetry sampled GPU clock, power and temperature, CPU scheduling, memory pressure and process I/O every 0.25 seconds. Candidate/baseline mean GPU clocks were 2838.66/2833.54 MHz during prefill and 2654.80/2666.77 MHz during generation; mean powers were 106.45/106.49 W and 105.61/105.25 W respectively. No sustained late-prefill clock collapse, process major faults, swapping or material CPU/I/O anomaly was observed. The candidate's approximately 0.45% lower generation clock is consistent with its approximately 0.45% lower rate, but this does not establish causality.
+
+**Profiler limitation:** ROCProfiler attachment was attempted only after each clean timing completed. Both attempts failed because the server had not been started with `ROCP_TOOL_ATTACH=1`. No cached diagnostic request ran and no kernel/API trace was obtained. The failed-attachment receipts are retained. Clean timing and system telemetry remain valid; there is no traced explanation of the earlier Sozo slowdown.
+
+### Earlier Sozo result retained
+
+The earlier candidate-fast comparison completed one baseline/patched pair at each length below. A duplicated follow-up pair was cancelled and its partial records are retained:
+
+| Work | Prefill change | Generation change | Output tokens |
+| --- | ---: | ---: | --- |
+| 8192 prompt / 128 output | -1.21% | -1.73% | Exact |
+| 65536 prompt / 128 output | -2.61% | -6.86% | Exact |
+
+That 64K pair failed the 2% loss screen and remains unexplained. An earlier general-path-only block also showed the unchanged baseline's 64K generation speed move from 16.56 to 15.40 tok/s. Clock and temperature traces were not captured for those runs. The later passing Ciru pair does not erase the failed Sozo gate, prove thermal throttling or establish parity under every machine condition.
+
+### Focused execution-cost checks
+
+A separate diagnostic used one patched and one baseline load with a synthetic 65536-cell cache and real full-model 1-, 7- and 512-token batches. The target context remained 262144, F16 KV and one slot; MTP was off. Attention/indexer values were zero-filled, sequence metadata was contiguous and recurrent state came from one token. Cache setup was excluded; one warmup and three small samples were taken per shape. These are operation costs, not real-conversation or served MTP rates.
+
+| Tokens in batch at 64K cache size | Median execution-time change (lower is faster) | Full-vocabulary output |
+| --- | ---: | --- |
+| 1 | +0.13% | Exact |
+| 7 | -0.15% | Exact |
+| 512 | -1.81% | Exact |
+
+All three cases passed their exact-logit and less-than-2% execution-cost gates. CPU mapping was faster in all 12 cases, with 24-58% less time for the 64K cases. The isolated GPU normalization comparison returned byte-identical outputs and added less than one microsecond per call at the 64K shape.
+
+**Qualification scope:** the sequence-isolation/copy/boundary checks, served recall, short MTP performance, later real 8K and 64K confirmations, and focused execution-cost checks support this isolated correctness release. The old Sozo slowdown and causal attribution remain unresolved. There is no filled-512K, multi-slot MTP, cross-layout bitwise invariance or new full task-quality-suite claim.
+
+The first general-path-only candidate was rejected: tokens matched, but pooled prefill losses were 3.40% at 8K and 4.16% at 64K, with noisy 64K decode timing. Its results remain in the evidence alongside the guarded candidate.
+
+## Build identity and evidence
+
+The NixOS candidate replaced only `libllama.so.0.3.0` in the retained runtime. Three affected translation units were rebuilt, with the guarded helper rebuilt afterward; the other 15 binaries were unchanged. The same candidate library was hash-verified on both Sozo and Ciru. The corresponding source also passed the earlier clean CPU build.
+
+A separate clean Ubuntu 24.04 / ROCm 10.0.0 build completed on Dunamis's Intel i9-14900KF in an isolated container with no GPU access and no inherited build objects. All 3566 input files were verified unchanged after compilation. `llama-server`, `llama-cli` and `llama-bench` built and passed ELF dependency/help checks. The existing tests passed: **66 QSA mapping/state cases** and **30 batch-allocation tests / 198 assertions**.
+
+The documented setup/build helper was used. Because the build machine was Intel, CPU compilation was explicitly configured with `GGML_NATIVE=OFF`, AVX2/F16C/FMA/BMI2 enabled and AVX512/VNNI disabled before the final build; this portability adjustment is recorded. No inference source changed.
+
+The resulting Ubuntu binaries were transferred by verified hashes to an Ubuntu 24.04 container on Ciru, using its NixOS host GPU driver and the matching ROCm 10 SDK. The 57-token coding prompt / 520-token greedy completion matched **all 520 reference token IDs**, with **674 drafted / 400 accepted**, at the production 262144 context allocation. **GPU device identity and all 15 binary hashes were verified.** This is a short active-prompt compatibility check, not an Ubuntu performance comparison or native-Ubuntu-host qualification.
+
+The first smoke verifier reported `gfx1151 identification not found` because it required that literal string in a verbosity-3 server log. The same built binary identifies ROCm0 as Radeon 8060S; the sole GPU KFD node reports `gfx_target_version=110501`, and the actual inference log records RDNA3.5 kernel execution. The reviewed result is PASS. The original verifier failure and raw evidence remain preserved; inference was not repeated to obtain that result.
+
+- NixOS candidate libllama SHA-256: `4a70e051c953617fe6a13d60444073dd2a5e7be1a9f707896841156939d3992c`.
+- Retained NixOS HIP library SHA-256: `f60bfe62b7dfab548c5a30dcf7728f47b85f142c39ca9ae7377cc340b71c19d4`.
+- Clean Ubuntu libllama SHA-256: `7d4f28ed7578e242a744b27e12d7a2a1d9256a1dbe6bb97a91b9b362be23066e`.
+- Clean Ubuntu HIP library SHA-256: `7955750301284227e36a558195fc4830ec1a521fed19e412159f3ac56a9629a9`.
+- [Final qualification decision](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/qualification/v2.0.1/FINAL-QUALIFICATION.json).
+- [Later Sozo 8K confirmation](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1/docs/qualification/v2.0.1/short-context-confirmation/).
+- [Ciru 64K timing, telemetry and independent review](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1/docs/qualification/v2.0.1/causal64k/).
+- [Clean Ubuntu build identities and CPU tests](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1/docs/qualification/v2.0.1/ubuntu24-build/).
+- [Reviewed Ubuntu GPU smoke receipt](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/qualification/v2.0.1/ubuntu24-gpu-smoke/REVIEWED-RESULT.json).
+- [All structured summaries and build/link identities](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1/docs/qualification/v2.0.1/).
+- [Raw performance records, frozen requests and diagnostic fixtures](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/releases/download/v2.0.1/qsa-v2.0.1-evidence.tar.gz).
+
+Build records distinguish hash-verified retained objects from one vendor-hash static archive omitted from the old sealed index; that archive was captured from its original link path and its identity is recorded. A retained executable's version string alone does not identify this library update. Use the tag/source manifest and library hashes.
+
+The separate HIP host-buffer proposal [#25863](https://github.com/ggml-org/llama.cpp/pull/25863) remains excluded after an unresolved MTP-output difference. It is not the fix for #27994. See [parallel validation history](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/PARALLEL_VALIDATION.md).

benchmarks/v2.0.1/git-source.json

@@ -0,0 +1,33 @@
+{
+  "version": "2.0.1",
+  "repository": "https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4",
+  "git_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "git_tree": "e060b9fce2a1ad3abee390fffe1c33ba02945e3b",
+  "intended_tag": "v2.0.1",
+  "base_public_commit": "145187f6aefb3ce8a39b32f06eeed3f85bea3e3b",
+  "tested_code_commit": "bf0bf4b795e112bb1b4af0101640b8fdb250365f",
+  "source_archive": {
+    "filename": "ciru-runtime-v2.0.1-source.tar.gz",
+    "sha256": "28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5",
+    "bytes": 36780030,
+    "prefix": "ciru-runtime-v2.0.1/",
+    "tracked_files_verified": 3593,
+    "verification": "Each Git blob content, regular-file mode and symlink target matches the named tar archive."
+  },
+  "evidence_archive": {
+    "filename": "qsa-v2.0.1-evidence.tar.gz",
+    "sha256": "a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66",
+    "bytes": 3644697,
+    "manifest_entries": 606
+  },
+  "source_manifest_sha256": "b93c6f1454f2a25d0fe346a8c2d1955f04a0d35c853cde6e4e446fbd3a955e5b",
+  "changed_core_sha256": {
+    "src/llama-memory-hybrid-idx.h": "4170f62854c1d8567b7d7fb15e673f4b6a242eaa1baab691944c1e3f4349e3ce",
+    "src/llama-kv-cells.h": "40c212f007a32047dad913dc20a090b8645556b3a2b32118e40cab2fb885ac09",
+    "src/llama-kv-cache.cpp": "d102bba0320cf786694119e1daa6681c1765bee6bde7559a7544143dce9e3280",
+    "src/llama-memory-hybrid-idx.cpp": "08971777e41badf654cf764f3b7e217304c19d94f1216c14e02abc5e14e1f3fd",
+    "src/models/qwen4exp.cpp": "6faf43e7693f6cebfb48a257f7a8fc6dab5cadeae88b712c048dac80d3aa268a"
+  },
+  "weights_changed": false,
+  "qualified_source_changes_after_build": "Release documentation and qualification receipts only; inference source/profile/build scripts unchanged."
+}

benchmarks/v2.0.1/publication/INDEPENDENT-PACKAGE-REVIEW.json

@@ -0,0 +1,316 @@
+{
+  "status": "PASS_STAGED_PACKAGE",
+  "reviewed_utc": "2026-09-07T23:17:29.235092+00:00",
+  "git_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "git_tree": "e060b9fce2a1ad3abee390fffe1c33ba02945e3b",
+  "source_archive_sha256": "28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5",
+  "checks": {
+    "git_commit_and_tree": {
+      "pass": true,
+      "details": {
+        "head": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+        "tree": "e060b9fce2a1ad3abee390fffe1c33ba02945e3b"
+      }
+    },
+    "worktree_clean": {
+      "pass": true,
+      "details": null
+    },
+    "github_staged_checksums": {
+      "pass": true,
+      "details": [
+        {
+          "path": "ciru-runtime-v2.0.1-source.tar.gz",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "qsa-v2.0.1-evidence.tar.gz",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "git-source.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "source-manifest.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "FINAL-QUALIFICATION.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "RELEASE-NOTES.md",
+          "exists": true,
+          "matches": true
+        }
+      ]
+    },
+    "huggingface_staged_checksums": {
+      "pass": true,
+      "details": [
+        {
+          "path": ".gitattributes",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "README.md",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "assets/qwen38-flash-ciru-strix-iu4.jpg",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "assets/v2-bf16-fidelity.png",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "assets/v2-context-sweep.png",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/FINAL-QUALIFICATION.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/QUALIFICATION.md",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/git-source.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/qsa-v2.0.1-evidence.tar.gz",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/source-manifest.json",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "checksums.sha256",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "profiles/strix-halo-production.env",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz",
+          "exists": true,
+          "matches": true
+        },
+        {
+          "path": "v2.0-checksums.sha256",
+          "exists": true,
+          "matches": true
+        }
+      ]
+    },
+    "source_archive_identity": {
+      "pass": true,
+      "details": null
+    },
+    "source_archive_matches_git_every_blob_and_mode": {
+      "pass": true,
+      "details": {
+        "files": 3593,
+        "expected_files": 3593,
+        "failures": [],
+        "missing": [],
+        "extra": []
+      }
+    },
+    "source_manifest_hash": {
+      "pass": true,
+      "details": null
+    },
+    "qualified_core_identity": {
+      "pass": true,
+      "details": [
+        {
+          "path": "src/llama-memory-hybrid-idx.h",
+          "matches_archive_and_worktree": true,
+          "matches_tested_commit": true
+        },
+        {
+          "path": "src/llama-kv-cells.h",
+          "matches_archive_and_worktree": true,
+          "matches_tested_commit": true
+        },
+        {
+          "path": "src/llama-kv-cache.cpp",
+          "matches_archive_and_worktree": true,
+          "matches_tested_commit": true
+        },
+        {
+          "path": "src/llama-memory-hybrid-idx.cpp",
+          "matches_archive_and_worktree": true,
+          "matches_tested_commit": true
+        },
+        {
+          "path": "src/models/qwen4exp.cpp",
+          "matches_archive_and_worktree": true,
+          "matches_tested_commit": true
+        }
+      ]
+    },
+    "changes_since_tested_code_are_docs_evidence_only": {
+      "pass": true,
+      "details": {
+        "changed_count": 58,
+        "unexpected": []
+      }
+    },
+    "github_hf_assets_identical": {
+      "pass": true,
+      "details": [
+        "ciru-runtime-v2.0.1-source.tar.gz",
+        "qsa-v2.0.1-evidence.tar.gz",
+        "git-source.json",
+        "source-manifest.json",
+        "FINAL-QUALIFICATION.json"
+      ]
+    },
+    "historical_hf_checksums_preserved": {
+      "pass": true,
+      "details": null
+    },
+    "production_profile_unchanged": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_archive_identity": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_member_manifest": {
+      "pass": true,
+      "details": {
+        "entries": 606,
+        "failed": []
+      }
+    },
+    "evidence_contains_short-context-confirmation/SUMMARY.json": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_contains_causal64k/SUMMARY.json": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_contains_causal64k/telemetry.jsonl": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_contains_ubuntu24-gpu-smoke/REVIEWED-RESULT.json": {
+      "pass": true,
+      "details": null
+    },
+    "evidence_contains_long-regression-fast/rows-so-far.json": {
+      "pass": true,
+      "details": null
+    },
+    "hf_qualification_source_references_present": {
+      "pass": true,
+      "details": [
+        {
+          "path": "docs/qualification/v2.0.1/FINAL-QUALIFICATION.json",
+          "present": true
+        },
+        {
+          "path": "docs/qualification/v2.0.1/short-context-confirmation/",
+          "present": true
+        },
+        {
+          "path": "docs/qualification/v2.0.1/causal64k/",
+          "present": true
+        },
+        {
+          "path": "docs/qualification/v2.0.1/ubuntu24-build/",
+          "present": true
+        },
+        {
+          "path": "docs/qualification/v2.0.1/ubuntu24-gpu-smoke/REVIEWED-RESULT.json",
+          "present": true
+        },
+        {
+          "path": "docs/qualification/v2.0.1/",
+          "present": true
+        },
+        {
+          "path": "docs/PARALLEL_VALIDATION.md",
+          "present": true
+        }
+      ]
+    },
+    "hf_new_release_local_references_present": {
+      "pass": true,
+      "details": [
+        {
+          "path": "runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz",
+          "present": true
+        },
+        {
+          "path": "benchmarks/v2.0.1/git-source.json",
+          "present": true
+        },
+        {
+          "path": "v2.0.1-checksums.sha256",
+          "present": true
+        }
+      ]
+    },
+    "qualification_material_limits_retained": {
+      "pass": true,
+      "details": {
+        "ROCP_TOOL_ATTACH=1": true,
+        "-2.61%": true,
+        "-6.86%": true,
+        "+2.1206%": true,
+        "-0.4488%": true,
+        "Ubuntu 24.04": true,
+        "Multi-slot MTP is still unsupported": true,
+        "#25863": true
+      }
+    },
+    "iu4_correction_present": {
+      "pass": true,
+      "details": null
+    },
+    "final_qualification_matches_repo": {
+      "pass": true,
+      "details": null
+    },
+    "local_tag_target": {
+      "pass": true,
+      "details": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e"
+    }
+  },
+  "blockers": [],
+  "notes": [
+    "No public upload or remote tag/link verification was attempted; new public links are publication targets.",
+    "Old unchanged HF v2.0 benchmark assets and model weights are intentionally absent from this incremental staging folder and were not downloaded or rehashed."
+  ],
+  "publication": "Not uploaded; independent local package verification only",
+  "writes": "Only this review receipt, outside source and release archives."
+}

benchmarks/v2.0.1/publication/PACKAGE-VERIFICATION.json

@@ -0,0 +1,26 @@
+{
+  "status": "PASS",
+  "publication": "Not uploaded; local staging only",
+  "git_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "git_tree": "e060b9fce2a1ad3abee390fffe1c33ba02945e3b",
+  "source_archive": {
+    "filename": "ciru-runtime-v2.0.1-source.tar.gz",
+    "sha256": "28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5",
+    "bytes": 36780030,
+    "prefix": "ciru-runtime-v2.0.1/",
+    "tracked_files_verified": 3593,
+    "verification": "Each Git blob content, regular-file mode and symlink target matches the named tar archive."
+  },
+  "evidence_archive": {
+    "filename": "qsa-v2.0.1-evidence.tar.gz",
+    "sha256": "a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66",
+    "bytes": 3644697,
+    "manifest_entries": 606
+  },
+  "github_hf_source_bytes_identical": true,
+  "historical_checksums_unchanged": true,
+  "full_and_runtime_only_checksum_checks": "PASS",
+  "source_code_hashes": "PASS",
+  "independent_smoke_checks": 24,
+  "evidence_originals_preserved": true
+}

benchmarks/v2.0.1/publication/READY.md

@@ -0,0 +1,34 @@
+# QSA v2.0.1 release readiness
+
+The focused QSA release is qualified and packaged. No public upload or production deployment has been performed. Ciru's GPU is released; Dunamis's build containers are stopped. Sozo's unrelated model service was left alone.
+
+| Check | Result |
+| --- | --- |
+| QSA mapping/state and canonical equivalence | 66 cases; 288 exact four-tensor comparisons |
+| Two-slot target-only serving | 16/16 recall requests across separate/unified KV, plus full-model isolation/copy controls |
+| Final 8K matched pair on Sozo | Prefill +4.70%, generation +1.20%; exact tokens/counts |
+| Final 64K matched pair on Ciru | Prefill +2.12%, generation -0.45%; exact tokens/counts |
+| Clean Ubuntu 24.04 / ROCm 10 build | Passed; 66 QSA cases and 30 batch tests / 198 assertions |
+| Ubuntu GPU smoke | All 520 tokens and 674 drafted / 400 accepted counts match; all 15 binary hashes verified |
+| Packaging | 3,593 Git paths verified, 606 evidence-manifest entries, full and runtime-only checksums pass |
+
+The clean build ran on Dunamis's CPU. Ciru performed one candidate/baseline 64K pair and one clean-build coding completion. No additional model run was made to correct the overly strict GPU-label verifier. Its original failure and independent adjudication are preserved.
+
+The earlier Sozo 64K loss remains unexplained. Final timing and hardware telemetry support this scoped release; the profiler attachment failed, so no kernel-level explanation was obtained. Multi-slot MTP remains unsupported. The HIP host-buffer patch, Kairic and HC-mix are excluded. No filled-512K, universal-agent, new full-quality-suite or native-Ubuntu-driver claim is made.
+
+## Source and delivery
+
+- Local branch: `codex/qsa-sequence-fix`.
+- Qualified source commit: `9ea2390a71ae9f3d1cab519bbe099eb4ee06380e`.
+- Git tree: `e060b9fce2a1ad3abee390fffe1c33ba02945e3b`.
+- Local annotated tag: `v2.0.1`, pointing to that commit; not pushed.
+- GitHub release assets: [release-ready/](release-ready/).
+- Hugging Face publication stage: [hf/](hf/).
+- [Package verification](PACKAGE-VERIFICATION.json), [qualification](../runtime/docs/QSA_BACKPORT_STATUS.md), [release notes](github-release.md), [community reply draft](COMMUNITY-REPLY.md).
+- [Independent package review](INDEPENDENT-PACKAGE-REVIEW.json): 25 checks passed, zero blockers.
+- Source archive SHA-256: `28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5`.
+- Evidence archive SHA-256: `a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66`.
+
+Only publication remains: push the source commits and new tag, create the GitHub release with these verified assets, upload the staged Hugging Face changes and pin its new v2.0.1 revision, then verify remote commit identities and downloaded source/checksum bytes. Preserve all v2.0 tags, archives, checksums and weights. The older prepared scripts are superseded by the final staging/packaging workflow and must not be rerun.
+
+No further GPU test is required for these unchanged, qualified source bytes. A later inference-code change or materially different combined patch requires its own qualification.

benchmarks/v2.0.1/publication/STAGED-CHECKSUMS.sha256

@@ -0,0 +1,14 @@
+c16fda8809ad10f0d25d2243f4f83e76c29b80e0ee1622a4357a6f7235df97b4  .gitattributes
+c81afc287d03d0375409dbe0d614bb1931b344aa20769a360853c9582dff0159  README.md
+3a46a7fd054305b2c1a931682015d93094312fb3b76c8ef83a2adb350b0f584c  assets/qwen38-flash-ciru-strix-iu4.jpg
+e260f72cba14e4c4e523794345f7de001444626240ab620d96dc44ee51beee90  assets/v2-bf16-fidelity.png
+88dd940fda0790ec6f474dc0a99019703114df253758ee14136c9686e93ea404  assets/v2-context-sweep.png
+2e1a7aa972d016ea70d3853d9328386717cb17f8802b02cbe9dae511484ca8a4  benchmarks/v2.0.1/FINAL-QUALIFICATION.json
+9d9e9d08ab443b616f2c693b08028f902cfee0fbb56a5ef997dd99183bf98fc0  benchmarks/v2.0.1/QUALIFICATION.md
+2f83bea64e1b302916bcc46f480c3b1040a3a5b5a9501cb73b11f63f32f96859  benchmarks/v2.0.1/git-source.json
+a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66  benchmarks/v2.0.1/qsa-v2.0.1-evidence.tar.gz
+b93c6f1454f2a25d0fe346a8c2d1955f04a0d35c853cde6e4e446fbd3a955e5b  benchmarks/v2.0.1/source-manifest.json
+cd1177aef73724230e04e53171321339177a7fcb6cd1292d6794cb9abfa98f54  checksums.sha256
+18a9f88143657e271ab7c6e8b8049219d1b267fcf7d993813a13eca5ed4ba82a  profiles/strix-halo-production.env
+28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5  runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz
+73bc5f5c4c99eac1670ecd2ecafdd0138829e3d88a7868178a837157d7c8d98e  v2.0-checksums.sha256

benchmarks/v2.0.1/publication/STAGED-MODEL-CARD.md

@@ -0,0 +1,300 @@
+---
+license: other
+license_name: qwen-community-1.0
+license_link: LICENSE
+base_model:
+  - Qwen/Qwen3.8-Flash-Next
+  - Qwen/Qwen3.8-Flash-Next-FP8
+base_model_relation: quantized
+library_name: llama.cpp
+pipeline_tag: text-generation
+inference: false
+tags:
+  - qwen
+  - qwen3.8
+  - qwen3.8-flash-next
+  - gguf
+  - llama.cpp
+  - amd
+  - rocm
+  - gfx1151
+  - ryzen-ai-max-395
+  - strix-halo
+  - mixture-of-experts
+  - iu4
+  - mtp
+  - speculative-decoding
+  - nvme
+  - ple
+  - long-context
+  - local-inference
+---
+
+[![Qwen3.8 Flash CIRU Strix IU4](assets/qwen38-flash-ciru-strix-iu4.jpg)](https://llm.ciru.ai/research)
+
+# Qwen3.8-Flash-CIRU-STRIX-IU4 · runtime v2.0.1
+
+Correctness, scoped performance and clean Ubuntu GPU qualification completed on 2026-09-07.
+
+**Mixed-precision Qwen3.8 for AMD Strix Halo, with ROCm 10 execution, NVMe-paged PLE, and depth-6 MTP.**
+
+**Runtime v2.0.1 fixes unified-KV conversation isolation and indexer cache copying.** The isolation/copy checks and later Sozo 8K / Ciru 64K serving gates passed. The earlier Sozo 64K slowdown remains unexplained; profiler attachment failed before any diagnostic request. [Qualification report](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/QSA_BACKPORT_STATUS.md). Model weights are unchanged; existing users only need to rebuild the runtime.
+
+The results below are the historical v2.0 measurements; v2.0.1 has its own bounded regression qualification.
+
+Version 2.0 brings together CIRU's retained ROCm 10 correctness and execution improvements. The measured strengths are **42.3 tok/s on the controlled non-thinking coding probe**, **382 tok/s cold prefill at 2K**, and **0.03045 mean forward KL to BF16** on the shared diagnostic panel. The target, MTP and PLE weights are unchanged.
+
+> [!IMPORTANT]
+> **Use the [v2.0.1 runtime archive](runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz) or matching [GitHub `v2.0.1` tag](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1).** This text-only package requires the custom CIRU runtime and all three `ple/` files. The `mtp/` draft enables speculation. Stock `llama.cpp` and Hugging Face hosted inference do not run this package correctly.
+>
+> **The target GGUF, MTP draft and PLE weights are unchanged.** Existing users only need the new runtime and profile. v2.0.1 identifies the runtime patch release; the production profile is unchanged.
+
+## Historical v2.0 release results
+
+| Result | CIRU v2.0 | Measurement |
+|---|---:|---|
+| Served non-thinking coding generation with MTP | **42.28–42.31 tok/s** | Two retained repetitions; 57-token prompt, 520 generated tokens, greedy coding profile |
+| Cold prefill at 2K | **382.03 tok/s** | 2,048 exact prompt tokens, MTP off |
+| Cold prefill at 128K | **232.95 tok/s** | 131,072 exact prompt tokens, MTP off |
+| Mean forward KL to BF16 | **0.03045** | 64 full-vocabulary distributions across four domains |
+| Observed-token PPL | **2.08401** | 60-token diagnostic slice; BF16 reference 2.01799 |
+
+The 42.3 tok/s result uses an explicitly closed thinking block and our retained fixed-six configuration: ROCm10, a 32,768-row draft shortlist, F16 target KV, Q8_0 draft KV, 16K context and the performance CPU governor. Both runs produced the same 520 output tokens. **It is a controlled coding-probe rate, not a HumanEval average or a claim that every workload runs at that speed.** [Recorded coding results](benchmarks/v2.0/controlled-coding-42tps.json).
+
+The comparison preserves each package's own recorded runtime settings. On the shared MTP-off sweep, CIRU has higher prefill from 2K through 128K: at 128K it is **26.4% faster than Agention/Laurent and 22.7% faster than Unsloth**. On the small BF16 panel, mean forward KL is **72.9% lower than Agention/Laurent and 82.6% lower than Unsloth**. These are fidelity and prefill measurements, not overall task-quality or generation-speed percentages.
+
+## Three-way comparison
+
+The three-way context sweep ran on **Sozo: Ryzen AI MAX+ 395 / Radeon 8060S, gfx1151, 128 GiB shared memory, NixOS**, with one model workload at a time. Ciru handled the new CIRU BF16 captures and the clean Ubuntu build/GPU smoke. The comparisons preserve each arm's recorded execution settings; they are package comparisons, not a controlled kernel-only experiment.
+
+| Arm | Target and runner | MTP implementation available in the package |
+|---|---|---|
+| CIRU v2.0 | Released mixed-format weights; locked RC2 ROCm 10 runtime | Fixed maximum 6, p-min 0, 32,768-row draft shortlist; target F16 KV, draft Q8_0 KV |
+| Agention / Laurent | [FP4 FAST model](https://huggingface.co/agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix-GGUF/tree/ad4c5717254a630ee0c5a8db5208eb1f8476e56c); [Laurent's Vulkan fork](https://github.com/LaurentZuijdwijk/llama.cpp/commit/5e085d123eead2e89b5c19f824fccb05727da6a2) | Publisher adaptive 2–4 with its FP4 draft; target Q8_0 KV |
+| Unsloth / recommended | [IQ4_XS model](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF/tree/38bb39ee97821de2c9009abb7e93950eec396e66); [Daniel Han Chen's MTP branch](https://github.com/danielhanchen/llama.cpp/commit/d1a92352cbd417fd840b4e765c0b82f5fe3d1d89), native Vulkan build | Unsloth [MTP README option 2](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF/blob/38bb39ee97821de2c9009abb7e93950eec396e66/MTP/README.md), maximum 2 and shared Q8_0 draft; publisher defaults |
+
+Only these three combinations are included. Exact model/shard hashes and runner commits are in [competitor artifacts](benchmarks/v2.0/competitor-artifacts.json); complete recorded settings are in [runtime recipes](benchmarks/v2.0/runtime-recipes.json).
+
+### BF16 numerical fidelity
+
+![BF16 fidelity diagnostic](assets/v2-bf16-fidelity.png)
+
+| Configuration | Mean forward KL ↓ | p95 KL ↓ | BF16 top-token agreement | Observed-token PPL ↓ |
+|---|---:|---:|---:|---:|
+| BF16 reference | 0 | 0 | 64/64 | 2.01799 |
+| CIRU ROCm10 v2.0 | **0.03045** | 0.15453 | 61/64 | 2.08401 |
+| Agention / Laurent | **0.11223** | 0.29765 | 60/64 | 2.38380 |
+| Unsloth / recommended | **0.17457** | 0.55722 | 61/64 | 2.42931 |
+
+This fixed diagnostic contains **64 full-vocabulary distributions and 60 observed next-token losses** across four short domain prefixes. MTP is off. CIRU uses F16 KV and flash-attention auto; the retained competitor captures use Q8_0 KV and flash attention on. CIRU's two independent model loads produced byte-identical logits.
+
+The result measures these implementations' fidelity on a small shared panel. It does not establish general task-quality superiority or full-corpus perplexity. The old CIRU/Q5 diagnostic used an earlier runtime and is superseded for the current three-way comparison.
+
+[Full statistics and hashes](benchmarks/v2.0/bf16-fidelity.json) · [CSV](benchmarks/v2.0/bf16-fidelity.csv)
+
+### MTP-off context sweeps
+
+![MTP-off context sweeps](assets/v2-context-sweep.png)
+
+Each cell is **prefill / generation tok/s**. All seven points use the same exact-token source fixture, a cold prompt, and 128 generated tokens.
+
+| Prompt tokens | CIRU ROCm10 v2.0 | Agention / Laurent | Unsloth / recommended |
+|---:|---:|---:|---:|
+| 512 | 306.53 / 22.55 | 357.63 / 25.93 | 235.03 / 24.54 |
+| 2,048 | 382.03 / 20.94 | 363.83 / 25.58 | 270.12 / 23.98 |
+| 8,192 | 370.41 / 19.19 | 325.59 / 24.95 | 274.78 / 22.69 |
+| 16,384 | 352.13 / 17.47 | 302.72 / 24.36 | 266.87 / 21.28 |
+| 32,768 | 321.33 / 14.32 | 273.95 / 22.77 | 254.68 / 18.79 |
+| 65,536 | 282.99 / 10.17 | 229.68 / 20.25 | 228.78 / 13.37 |
+| 131,072 | 232.95 / 6.75 | 184.33 / 17.86 | 189.79 / 9.98 |
+
+Server context is 262,144, sampling is greedy with seed 1234, EOS is ignored for the 128-token measurement, and a 512+32 warmup is excluded. Each row verifies exact prompt/output counts and zero drafted/accepted tokens.
+
+CIRU leads prefill at 2K and above. **Both competitor arms have higher MTP-off generation rates across this sweep.** These are target-only context measurements, separate from the native MTP panel. Unsloth's target-only sweep uses Q8_0 KV, 16 threads, batch 2048/microbatch 512 and explicit CPU PLE placement with lazy mode off; its native MTP panel uses publisher defaults. CIRU uses F16 KV, 8 threads, and its 4 GiB PLE cache. See the recorded recipes for all differences.
+
+[Full sweep CSV](benchmarks/v2.0/mtp-off-context-sweep.csv) includes TTFP and memory. [Structured results](benchmarks/v2.0/comparison.json) preserve idle/peak/delta RAM, GTT and VRAM counters. RAM is whole-system usage; these overlapping shared-memory counters must not be summed.
+
+
+
+## What changed in v2.0
+
+The locked candidate includes the retained Sozo quality-lane and Ciru speed-lane fixes:
+
+- Correct attention descriptor strides, GPU admission for long QSA top-k, and restored-KV fast-path admission.
+- Seven-column Q5_K weight reuse, the GPU MTP top-10 selector and a 32,768-row draft-output shortlist.
+- Decode-entry and internal-microbatch synchronization repairs.
+- Fresh HIP graph recapture, plus a launcher working-directory fix.
+- The complete ROCm SDK installation helpers from v1.1.1, with the updated depth-6 production profile.
+
+The previous allocator-lifetime and cached-prefix/MTP state fixes remain included. The adaptive runner and broad expert-reuse/fusion experiments are not enabled in this release. No weights were requantized or retrained.
+
+The v2.0 core source matched **`qwen38-ciru-rocm10-20260905-rc2`**. v2.0.1 adds the QSA backport and guarded host path described in its qualification report. Portable build helpers are based on public v1.1.1 (`764ee491`); the retained core patch is based on v1.1 (`baba5e06`). [Provenance and runtime hashes](benchmarks/v2.0/ciru-provenance.json) record the exact inclusion evidence. The additional long-QSA/restored-KV admission code does not by itself establish a measured production cache-recovery speedup.
+
+## Download and build
+
+The model artifacts total **135,962,881,135 bytes (126.625 GiB)**, excluding runtime source and reports. Target storage is 73.945 GiB, the MTP draft is 3.852 GiB, and the mandatory PLE payload is 48.828 GiB. The target and all three `ple/` files are required; the MTP draft is optional only when speculation is disabled.
+
+On Ubuntu/Debian, install `python3-venv`, then download the package:
+
+```bash
+sudo apt update
+sudo apt install -y python3-venv
+python3 -m venv .venv-hf
+.venv-hf/bin/python -m pip install -U huggingface_hub
+. .venv-hf/bin/activate
+hf download jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 \
+  --revision v2.0.1 --local-dir ./Qwen3.8-Flash-CIRU-STRIX-IU4
+cd Qwen3.8-Flash-CIRU-STRIX-IU4
+sha256sum -c checksums.sha256
+sha256sum -c v2.0.1-checksums.sha256
+tar -xzf runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz
+cd ciru-runtime-v2.0.1
+./scripts/ciru/setup-linux-amd.sh --install-host-deps
+```
+
+The helper installs host build tools and a private, complete **ROCm 10.0.0** SDK with gfx1151 device libraries. It builds into `build-gfx1151-sdk/`. Keep `.venv-rocm/` for runtime libraries. The host must already provide a compatible AMD GPU driver and access to `/dev/kfd` and its render node.
+
+**Existing users:** keep your model directory. Download only the runtime archive, updated profile and release checksums into a separate directory, then rebuild:
+
+```bash
+hf download jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 \
+  --revision v2.0.1 --local-dir ./ciru-v2.0.1-update \
+  --include "runtime/v2.0.1/*" "profiles/*" "v2.0.1-checksums.sha256"
+cd ciru-v2.0.1-update
+sha256sum -c v2.0.1-checksums.sha256 --ignore-missing
+tar -xzf runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz
+cd ciru-runtime-v2.0.1
+./scripts/ciru/setup-linux-amd.sh --install-host-deps
+```
+
+The source archive includes build/run instructions, provenance, licenses, tests and qualification summaries. The matching [GitHub `v2.0.1` tag](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1) has identical file contents and Git modes. [Release downloads](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/releases/tag/v2.0.1) · [Diff from v2.0](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/compare/v2.0...v2.0.1) · [Source identity](benchmarks/v2.0.1/git-source.json).
+
+As an alternative to extracting the archive, with Git installed:
+
+```bash
+git clone --branch v2.0.1 --single-branch https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4.git ciru-runtime-v2.0.1
+cd ciru-runtime-v2.0.1
+./scripts/ciru/setup-linux-amd.sh --install-host-deps
+```
+
+The original v2.0 archive and tags remain unchanged. v2.0.1 is a new source release with the QSA fix and the early multi-slot MTP launcher check.
+
+v2.0.1 passed a clean Ubuntu 24.04 / ROCm 10 build, ELF checks, 66 QSA mapping/state cases and 30 batch-allocation tests. Its 57-token coding prompt / 520-token completion matched every reference token and the 674 drafted / 400 accepted counts at 262144 context. GPU device identity and all 15 binary hashes were verified; the original overly strict device-label verifier failure is preserved in the qualification report. This uses Ubuntu container userspace on Ciru’s NixOS host GPU driver; the separate performance comparisons use NixOS binaries. The later Sozo 8K pair measured +4.6990% prefill / +1.2032% generation; the later Ciru 64K pair measured +2.1206% / -0.4488%, with matching output/work in each pair. The older failed Sozo 64K result and failed profiler attachment remain in the qualification report. [Build qualification](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/BUILD_LINUX.md).
+
+## Run the production profile
+
+From the cloned or extracted runtime directory:
+
+```bash
+BUILD_DIR="$PWD/build-gfx1151-sdk" \
+  MODEL_DIR=/absolute/path/to/Qwen3.8-Flash-CIRU-STRIX-IU4 \
+  ./scripts/ciru/run-server.sh
+```
+
+The profile uses one slot, a 262,144-token context, batch 2048/microbatch 512, 8 target threads, F16 target KV, Q8_0 draft KV, flash attention on, maximum MTP depth 6, p-min 0 and p-split 0.10. It enables a separate 4 GiB PLE cache with 16 workers, an 8 GiB prompt cache, idle-slot caching and context checkpoints. The server binds to `127.0.0.1:8080`.
+
+### Confirm the MTP profile
+
+The launcher exports `profiles/strix-halo-production.env`, including `CIRU_MTP_SHORTLIST=32768` and `CIRU_MTP_TOPK10=1`. Compiling v2.0 and setting `--spec-draft-n-max 6` alone does not enable those optimizations. With MTP enabled, confirm this startup line:
+
+```text
+CIRU MTP shortlist enabled: 32768 / 248320 vocabulary rows; full target verification retained
+```
+
+For direct `llama-server` launches, use the [complete exported environment and flags](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md#2-launch-the-audited-public-profile).
+
+### Choose a draft depth
+
+**Depth 6 is the released default; the fastest depth depends on the workload.** When later draft tokens are frequently rejected, depth 3 may be faster, including for some short interactive completions. Restart with this command to change only the maximum depth:
+
+```bash
+MTP_DEPTH=3 BUILD_DIR="$PWD/build-gfx1151-sdk" \
+  MODEL_DIR=/absolute/path/to/Qwen3.8-Flash-CIRU-STRIX-IU4 \
+  ./scripts/ciru/run-server.sh
+```
+
+The controlled coding result above does not establish that six is optimal for short chat, long contexts or batched serving. The context sweeps have MTP off and do not compare draft depths. No adaptive depth selection is enabled. Single sampled completions are useful directional feedback, but need matched settings and repeated measurements to support a general speed claim. See [comparison details](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/main/docs/RUNNING.md#confirm-the-mtp-profile-and-choose-a-draft-depth).
+
+### Send a chat request
+
+```bash
+curl http://127.0.0.1:8080/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "Qwen3.8-Flash-CIRU-STRIX-IU4",
+    "messages": [{"role": "user", "content": "Hello"}],
+    "temperature": 1.0,
+    "top_p": 0.95,
+    "top_k": 20,
+    "min_p": 0,
+    "cache_prompt": true
+  }'
+```
+
+For target-only serving, set `ENABLE_MTP=0` before launching. This omits the draft flags. Use the launcher option rather than appending `--spec-type none` to an already speculative command.
+
+Thinking defaults are temperature 1.0, top-p 0.95, top-k 20, min-p 0. For non-thinking mode, send `"chat_template_kwargs":{"enable_thinking":false}` with temperature 0.7, top-p 0.8 and presence penalty 1.5. Normal chat can retain prefix caching; the cold benchmark settings are not the production preset.
+
+`min_p=0` explicitly follows [Qwen's model-specific recommendation](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/f5d08274bafd880402bd16f5e3e6c514136ec06c/README.md#best-practices); v1.1 could inherit llama.cpp's `0.05`. This disables an additional probability filter and can change sampled output. We have not established that either value gives better answer quality in a controlled A/B test. Set `MIN_P=0.05` with the launcher to select the previous inherited value; request-level settings can override the server default.
+
+### Parallel agents and unified KV
+
+**The MTP shortlist still supports exactly one slot in v2.0.1.** The launcher rejects multi-slot MTP before model load. For two slots, disable MTP explicitly:
+
+```bash
+ENABLE_MTP=0 PARALLEL_SLOTS=2 CONTEXT_SIZE=524288 \
+  BUILD_DIR="$PWD/build-gfx1151-sdk" MODEL_DIR=/absolute/path/to/model \
+  ./scripts/ciru/run-server.sh --no-kv-unified
+```
+
+For unified KV on **v2.0.1**, replace `--no-kv-unified` with `--kv-unified`. Both modes passed the bounded concurrency, isolation and recall checks. Rebuild the v2.0.1 source before using unified multi-slot serving; the original v2.0 runtime is missing this fix.
+
+With separate KV, `524288` is the total context allocation: **262144 tokens per slot**, not 512K per agent. Unified KV shares the pool; this runtime still caps each slot at the model's 262144-token training context. Inspect `/slots` for actual limits.
+
+v2.0.1 includes the relevant QSA fix from [upstream #27941](https://github.com/ggml-org/llama.cpp/pull/27941) for [#27994](https://github.com/ggml-org/llama.cpp/issues/27994). QSA blocks now separate sequence sets, and pending copies update indexer keys. Full-model simultaneous-sequence tests retained exact logits when the other conversation changed. Sixteen served recall requests passed across split/unified KV with 13062-token initial prompts, concurrent requests, followups and slot reuse.
+
+These checks do not validate a filled 512K cache or every agent workload. The earlier v2.0 short-marker smoke missed the QSA issue; it should not be used as evidence that v2.0 is protected. The separate HIP host-buffer proposal #25863 remains excluded after its MTP regression comparison.
+
+[v2.0.1 qualification](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/QSA_BACKPORT_STATUS.md) · [Parallel validation history](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0.1/docs/PARALLEL_VALIDATION.md).
+
+## Historical task-quality results on the unchanged weights
+
+These earlier H96/depth-1 results were **not rerun as full suites on v2.0 or v2.0.1**. v2.0 evidence comprises the bounded BF16 panel, controlled coding probe and context sweeps above. v2.0.1 adds the scoped QSA correctness/build and performance regression checks.
+
+| Benchmark | Earlier result | Scope |
+|---|---:|---|
+| HumanEval | 160/164 (97.56%) | Full local-custom chat run, first sample |
+| HumanEval+ | 155/164 (94.51%) | Full EvalPlus 0.1.10 run |
+| ARC-Challenge | 1,143/1,172 (97.53%) | Full EvalScope dataset |
+| ToolEval Standard | 115/138 (83.33%) | 69 local-custom cases |
+| ToolEval Hard | 23/30 (76.67%) | 15 local-custom cases |
+
+Historical methodology and additional sampled results remain available in the [previous model card](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/9327d8c4b796d983d97870f0908e8667c0cbab2d/README.md). Local-custom or sampled scores are not canonical leaderboard submissions.
+
+## Storage and hardware
+
+The target contains 1,223 tensors: 144 routed-expert Q4_1 tensors, 328 Q5_K, 290 Q8_0, 48 Q5_1, 25 BF16 and 388 F32. The existing IU4 name is retained for continuity. **The standard launcher does not activate the separate native IU4/E3 bank path.** Its ordinary Q4_1 matrix path expands packed values into byte lanes and uses IU8 WMMA; other shapes can use vector kernels. These are standard GGUF tensor types. This corrects the execution description without changing the weights or the historical benchmark path. The PLE payload preserves exact FP8 E4M3 weights with one BF16 scale.
+
+The intended machine is a Ryzen AI MAX+ 395 / Radeon 8060S with 128 GiB unified memory and fast NVMe. Allow at least 160 GiB for model files and verification, plus space for the source, SDK and build. Linux provides the measured P16/O_DIRECT pager path. The configured context is 262,144; cold sweep coverage stops at 131,072.
+
+[File tree](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/tree/v2.0.1) · [Weight checksums](checksums.sha256) · [Current runtime/report checksums](v2.0.1-checksums.sha256) · [Historical v2.0 checksums](https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v2.0/v2.0-checksums.sha256) · [Measurement methodology](benchmarks/v2.0/METHODOLOGY.md)
+
+## Lineage, license and credit
+
+- Text lineage: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/tree/f5d08274bafd880402bd16f5e3e6c514136ec06c).
+- PLE lineage: [Qwen/Qwen3.8-Flash-Next-FP8](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8/tree/bcd9f01ddc9cff2316eb84281bebcd5b058bddce).
+- Runtime base: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/commit/f5e85d43a048f3d5adefb4c5e29867d8077fba62).
+
+Model artifacts use the included **Qwen Community License 1.0**; runtime code retains its MIT license and component notices. Credits include Qwen, ggml-org, Ryan Monsurate's MTP integration, AMD's ROCm ecosystem, and contributors recorded in the source notices. Thanks also to Laurent Zuijdwijk, Agention AI, Unsloth and Daniel Han Chen for the publicly available comparison models and runtimes, and OpenAI's HumanEval authors for the benchmark tasks.
+
+CIRU is an independent community research project. AMD and Qwen marks do not imply sponsorship or endorsement.
+
+```bibtex
+@software{ciru_qwen38_flash_strix_iu4_2026,
+  title   = {Qwen3.8-Flash-CIRU-STRIX-IU4},
+  author  = {{CIRU}},
+  year    = {2026},
+  version = {2.0},
+  url     = {https://huggingface.co/jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4}
+}
+```

benchmarks/v2.0.1/qsa-v2.0.1-evidence.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66
+size 3644697

benchmarks/v3.0.0/COMPARISON.md

@@ -0,0 +1,128 @@
+# CIRU Strix runtime v3.0.0 qualification
+
+V3 combines the qualified QSA fixes from the prior READY package with the portable prefill and decode improvements from the retained hybrid checkpoint. Target, PLE and MTP weights are unchanged. The general v3 profile retains MTP 6 and uses b1024/u1024; Halo uses its own best screened stock profile, MTP 3 on Vulkan.
+
+## Measured serving performance
+
+Ciru: Ryzen AI Max+ 395, gfx1151, 128 GB shared memory, NixOS. One model workload ran at a time. These cold coding requests use exactly matching chat-formatted input token IDs, 128 generated tokens, one slot and 262144 context capacity. EOS is honored; every reported request produced all 128 tokens. The sampler is nonthinking, temperature 0.7, top-p 0.8, top-k 20, min-p 0, presence penalty 1.5, repeat penalty 1, frequency penalty 0 and seed 123.
+
+| Input tokens | Profile | Prompt tok/s | Generation tok/s | First streamed piece (s) | Whole request (s) | Rows |
+| ---: | --- | ---: | ---: | ---: | ---: | ---: |
+| 4,096 | Previous CIRU | 392.00 | 22.52 | 10.70 | 16.34 | 2 |
+| 4,096 | CIRU v3 | 455.65 | 24.60 | 9.25 | 14.41 | 6 |
+| 4,096 | Halo | 381.49 | 35.30 | 11.09 | 14.69 | 4 |
+| 65,536 | Previous CIRU | 284.49 | 13.33 | 230.46 | 239.99 | 1 |
+| 65,536 | CIRU v3 | 369.81 | 24.22 | 177.32 | 182.57 | 3 |
+| 65,536 | Halo | 263.42 | 23.28 | 248.91 | 254.37 | 2 |
+
+| Input tokens | Comparison | Prompt rate change | Generation rate change | Request time reduction |
+| ---: | --- | ---: | ---: | ---: |
+| 4,096 | v3 vs Previous CIRU | +16.24% | +9.25% | +11.80% |
+| 65,536 | v3 vs Previous CIRU | +29.99% | +81.66% | +23.93% |
+| 4,096 | v3 vs Halo | +19.44% | -30.32% | +1.93% |
+| 65,536 | v3 vs Halo | +40.39% | +4.04% | +28.23% |
+
+Prompt processing and whole-request latency are distinct from generation speed. The table exposes each result; an overall request-time advantage is not a claim that CIRU wins every generation-rate measurement. The fresh MTP 6 confirmation is close to Halo at 4K, so small aggregate short-request differences should not be treated as a robust default-profile win. The qualified MTP 2 option provides the clearer short-request latency benefit on this fixture; the long-context improvement is much larger. Different CIRU/Halo weights and runtime profiles make the external result a serving-package comparison, not a runtime-only or quantization-quality comparison. The previous/current CIRU pair uses identical model weights but each release's own profile, so it measures the combined runtime/profile update.
+
+The table includes three clean v3 MTP 6 loads: the initial C1/C2 pair and a fresh final confirmation. Halo MTP 3 has two selected-profile loads. Between the initial and final v3 runs, the stock MTP configuration screens, selected Halo A1, previous CIRU and optional v3 MTP 2 checks ran. The fresh v3 MTP 6 confirmation was followed by Halo A2 and CIRU full capacity. Each load has one excluded 512/16 warmup. Each speed position has 4K/128, 64K/128, then a second 4K/128 request. A1, previous CIRU and C1 then run quality. No build, weight download or profiling runs alongside measured serving. This is a small repeated experiment, not a confidence interval or every-workload ranking.
+
+Aggregated PP divides total prompt tokens by total native prompt time. TG follows the native N-1 convention: total 127 timed output tokens per request divided by total native generation time. Latencies are arithmetic means from the official API harness; first-piece time is its first streamed content-field event. Loading, warmups, quality request speeds and instrumented requests are excluded. Raw row values, ranges, MTP counts and exact input hashes are in comparison.json.
+
+## Configuration selection and execution evidence
+
+The following are single 4K/128 selection screens using the published sampler, except the batch screen, which used greedy sampling. They selected finalists; the repeated final measurements above carry the release claim.
+
+| Runtime | Maximum draft depth | Prompt tok/s | Generation tok/s | Drafted / accepted |
+| --- | ---: | ---: | ---: | ---: |
+| CIRU | 2 | 454.94 | 28.88 | 112 / 71 |
+| CIRU | 3 | 451.18 | 27.94 | 159 / 73 |
+| CIRU | 4 | 436.93 | 26.04 | 200 / 77 |
+| CIRU | 6, initial repeated profile | about 459 | about 24.7 | see initial rows |
+| Halo | 2, initial repeated profile | about 383 | about 31.4 | see initial rows |
+| Halo | 3 | 383.66 | 35.37 | 132 / 82 |
+| Halo | 4 | 381.41 | 30.04 | 165 / 85 |
+| Halo | 6 | 376.20 | 25.91 | 214 / 90 |
+| Halo | 6, native adaptive | 379.69 | 29.31 | 145 / 83 |
+
+Higher maximum depth was tested on Halo without source changes. Its native maximum 3 won this screen; using the largest supported value would have made this workload slower. The CIRU batch/microbatch 512, 1024, 1536, 2048 screen measured PP 408.31, 451.19, 449.25, 390.17 tok/s respectively. 1024 was selected over the close 1536 result for the smaller workspace and qualified at 64K and full capacity. The MTP 2 setting is retained as an option for low-acceptance long requests, rather than replacing the general MTP 6 default. The high-acceptance coding counterexample below prevented that default change. These selections do not establish a universal optimum.
+
+A native ROCm fresh-launch trace confirmed radix TOP_K, indexed decode attention and tiny F32 gather dispatches, alongside the device-resident MTP chain. The trace captured 204 radix-selector, 132 indexed-attention and 132 tiny-gather dispatches; [kernel-path evidence](https://github.com/ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4/blob/v3.0.0/docs/qualification/v3.0.0/kernel-path-summary.json) records the counts and diagnostic timing scope. The trace's served rates are excluded. A focused extension of existing Q5_K weight reuse to three verification columns passed 17 actual GPU reference cases but measured 27.47 TG versus 28.88 for the existing path. It was reverted and parked. The release binaries match the pre-candidate hashes; no benefit from this rejected change is claimed.
+
+The earlier Halo MTP 2 mirrored 64K runs varied from 13.28 to 20.61 TG despite identical output IDs and 111 drafted/71 accepted tokens. PP was 265.42 and 264.01. The cause was not established; those results are retained as preliminary evidence and are not substituted for the selected MTP 3 comparison.
+
+## Quality and correctness
+
+| Profile | HumanEval base | EvalPlus extended tests | Two-key recall, original + cached | Exact cached token replay |
+| --- | ---: | ---: | --- | --- |
+| Previous CIRU | 20/20 | 20/20 | 4/4 at ~8K and ~64K | Both lengths |
+| CIRU v3 | 20/20 | 20/20 | 4/4 at ~8K and ~64K | Both lengths |
+| Halo | 20/20 | 20/20 | 4/4 at ~8K and ~64K | Both lengths |
+
+## MTP depth depends on the workload
+
+The same 20 short coding requests provide a counterexample to promoting MTP 2 as a universal default. Their generation work has high draft acceptance. The previous CIRU and v3 MTP 6 runs produced identical token IDs on all 20 tasks; MTP 2 and Halo have different output lengths, shown below. These are diagnostic native compute rates over the bounded quality panel, separate from the fixed-output speed experiment.
+
+| Profile | Generated tokens | Prompt tok/s | Generation tok/s | Sum of request times (s) |
+| --- | ---: | ---: | ---: | ---: |
+| Previous CIRU | 3179 | 148.53 | 53.33 | 75.49 |
+| CIRU v3 MTP 6 | 3179 | 219.51 | 53.24 | 70.67 |
+| CIRU v3 MTP 2 | 3212 | 226.25 | 39.63 | 91.57 |
+| Halo MTP 3 | 3241 | 176.10 | 49.48 | 79.15 |
+
+MTP 6 retains high-acceptance coding throughput while v3 reduces prompt time. MTP 2 improves the lower-acceptance 4K/64K coding fixture but is slower on these short tasks, so it is an explicit option instead of the general default. Halo MTP 3 is the selected-profile result here; this table is not a claim that it is the optimal Halo depth on every short coding task.
+
+| Input tokens | Optional CIRU MTP 2 prompt tok/s | Generation tok/s | Whole request (s) |
+| ---: | ---: | ---: | ---: |
+| 4,096 | 453.09 | 29.39 | 13.62 |
+| 65,536 | 373.08 | 24.88 | 180.87 |
+
+The panel is canonical HumanEval tasks 0–19, EvalPlus v0.1.10 base and extended tests, first sample only, no retries, 4096-token cap and truncations counted as failure. Generated code is tested inside a filesystem/network sandbox. This bounded nonthinking coding/recall panel does not establish broad model equality, thinking-mode quality, benchmark rankings, tool-use reliability or filled-context accuracy.
+
+The final inference sources passed 69 QSA mapping/state/guard cases with flags on and off, 33 ROCm TOP_K/GET_ROWS CPU-reference cases and 30 batch allocator tests with 198 assertions. Cases include ties, selection/gather admission boundaries and independent sequences with equal cell offsets. An earlier test fixture requested unsupported partial KV copying; the corrected full-copy fixture passes, and the failed attempt is preserved.
+
+The four-prefix same-weight numerical diagnostic matched 15,892,480 F32 logits byte-for-byte against the previous runner. It uses four 64-token prefixes, 64 full-vocabulary vectors, context 512, one-token microbatch and MTP off. It does not exercise long-prefill ordering. Radix selection can change threshold-tie membership and selected-list order; general bitwise equivalence is not claimed.
+
+## Full-capacity serving
+
+The final CIRU profile completed 261,888 real input tokens plus 128 generated tokens within 262,144 capacity: **257.44 prompt tok/s, 18.00 generation tok/s, 1017.38 s first piece and 1024.44 s whole request**. It was a cold request with EOS honored. This verifies serving capacity on this host, not full-context task accuracy. No filled 256K Halo comparison is claimed.
+
+## Included runtime and prior package
+
+The base `9ea2390a71ae9f3d1cab519bbe099eb4ee06380e` already includes the READY package's QSA isolation/copy repairs. V3 adds wider aligned QSA prefill, J32 expert tiling on supported prefill shapes, parallel radix cell selection, indexed F16 decode attention, guarded direct PLE previous-token lookup, F32 derived-history reuse and packed tiny F32 gathers. Derived-cache validity includes sequence identity. Each new environment switch accepts literal 0. Draft-window code is included but disabled and unqualified when enabled.
+
+External-GPU ownership, peer transfer, hot/cold expert placement, scheduler overlap, compact masks and the parked bulk-prefill pooling extension are excluded. The checkpoint's contiguous expert-ID fix belongs to that excluded hybrid routing path. The derived F32 cache adds 192 MiB at 256K over the prior F16 allocation: 12 layers × 128 values × 65,536 blocks × 2 extra bytes. The width 128 is verified from the actual GGUF indexer-normalization tensors.
+
+The original READY.md, package verification, independent package review, source/evidence archives and their checksums are preserved in the prior-package archive and historical qualification directory. v2.0.1 was qualified locally but not separately published as a tag; its fixes ship in v3.0.0. Existing public v2.0 tags, weights and weight checksums remain unchanged.
+
+A clean initial same-profile, same-weight greedy 64K/128 pair measured previous CIRU 290.34 PP / 13.95 TG and v3 341.57 PP / 25.50 TG (+17.64% / +82.73%). All 128 output IDs and 225 drafted/89 accepted counts matched. This one-sequence pair predates the sequence-key guard and uses b2048/u512 and MTP 6; it is separate from the final release-profile comparison.
+
+The initial 4K raw-record continuation ended after four text tokens; the then-enabled ignore_eos forced repeated end tokens. That unsuitable speed row is preserved and excluded. All reported final speed inputs use a real coding request and honor EOS. Initial quality runs overlapping CPU build/download activity and the profiler rates are also excluded from clean performance claims.
+
+Isolated 262144-column TOP_K measured 12,062.33 us to 1,792.08 us for 4 rows (6.73x), and 288,363.00 us to 48,854.72 us for 1536 rows (5.90x), versus heap selection with warp scan. Both use the 16-node graph duplication cap. These are kernel measurements, not served speed multipliers.
+
+## Halo compatibility and settings
+
+Halo source is pristine commit `5f851647fe5ed795dfd6c0a3fba543114879e874`, using its recommended Vulkan backend. The target is Unsloth UD-Q4_K_XL at revision `38bb39ee97821de2c9009abb7e93950eec396e66`, all four shards verified against LFS hashes. It uses full GPU offload, documented --ngram-on-disk, selected maximum MTP 3, and native KV, batch, thread, fitting and cache defaults. Only supported command-line MTP settings were screened.
+
+Unsloth's 2.79 GB shared Q8 head intentionally omits tensors a supporting runtime borrows from its main model. The pinned Halo loader requires token_embd.weight and fails even after loading the main model. Unsloth's self-contained head also fails for missing output_hc_norm.weight. The comparison uses the already published EasiiX Strix Q8 head at revision `6f7900648b1c6b14f067a182c640e47971e9ab35`, SHA256 `9db03a687670608286e99b563fcc86d0ee76c8dd863f64b2afc0b54eb0eb975d`. No Halo source, weights or adapters were modified. Both failed compatibility attempts are preserved.
+
+CIRU retains its 79.40 GB target, 52.43 GB PLE payload and 4.14 GB Q8 draft, ROCm 10, F16 target KV, Q8 draft KV, 32,768-row draft shortlist, eight CPU threads, one slot and full target verification. The general maximum MTP depth is 6; MTP_DEPTH=2 selects the tested long-request option. The optional draft attention window remains off.
+
+Setup sources: [Halo README](https://github.com/halo-box/strix-llama.cpp/blob/5f851647fe5ed795dfd6c0a3fba543114879e874/README.md), [Unsloth guide](https://unsloth.ai/docs/models/qwen3.8-next), [Unsloth MTP card](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF/blob/38bb39ee97821de2c9009abb7e93950eec396e66/MTP/README.md), [EasiiX head](https://huggingface.co/EasiiX/Qwen3.8-Flash-Next-MTP-Strix-Halo-GGUF/blob/6f7900648b1c6b14f067a182c640e47971e9ab35/README.md).
+
+## Memory, platform and reproducibility
+
+| Input tokens | Profile | Mean RAM before request (GiB) | Peak system RAM (GiB) | Max RAM increase (GiB) | Peak GTT (GiB) | Peak VRAM (GiB) |
+| ---: | --- | ---: | ---: | ---: | ---: | ---: |
+| 4,096 | Previous CIRU | 100.73 | 103.43 | 2.15 | 88.58 | 0.52 |
+| 4,096 | CIRU v3 | 103.06 | 106.02 | 2.43 | 90.29 | 0.53 |
+| 4,096 | Halo | 102.43 | 104.88 | 2.77 | 93.08 | 1.85 |
+| 65,536 | Previous CIRU | 100.59 | 101.32 | 0.72 | 88.58 | 0.52 |
+| 65,536 | CIRU v3 | 102.91 | 103.66 | 0.74 | 90.29 | 0.52 |
+| 65,536 | Halo | 103.14 | 103.72 | 0.60 | 93.17 | 1.99 |
+
+RAM, GTT and VRAM overlap on this unified-memory APU and must not be added. These are sampled host/device telemetry values, not model-file sizes.
+
+New GPU qualification is NixOS/ROCm 10/gfx1151. Historical clean Ubuntu qualification belongs to v2.0.1; no new v3 Ubuntu result is claimed. Source build instructions are retained. The optional binary payload is the tested NixOS server, bench, focused tests and shared libraries; it requires the recorded Nix store and SDK paths. No tested CLI binary is claimed. Vision and simultaneous multi-slot MTP remain outside this qualification.
+
+Raw requests, SSE, output IDs, samplers, memory samples, model/build identities, quality scoring artifacts, failures, official speed row hashes and configuration screens are preserved in the evidence archive. Source archives are checked against the release Git tree including modes and symlinks. The prior READY package is included separately with its original checksums. The main service remains in its original inactive, enabled, unmasked state.

benchmarks/v3.0.0/FINAL-QUALIFICATION.json

@@ -0,0 +1,245 @@
+{
+  "version": "3.0.0",
+  "source": {
+    "version": "3.0.0",
+    "tag": "v3.0.0",
+    "commit": "e2b584751f89e02334304365edfe4c5404fd9e5c",
+    "tree": "106fbbf40b85acf8ad311ac0a5fdf801df3c23f3",
+    "base_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+    "archive": "ciru-runtime-v3.0.0-source.tar.gz",
+    "archive_sha256": "5680641c5f38eb5e8fe23f9af73bc4347b76c670e2425e1f6a2bbb9ebef1723a",
+    "file_count": 3615,
+    "manifest": "source-manifest.json",
+    "verification": "Every file content, executable mode and symlink target matches the release Git tree; named archive bytes are verified independently."
+  },
+  "source_patch_sha256": "47711a875f7b5a56395ca4825f8ebfa80a36c0f2987bfe02e4f4f531a1d8cbf6",
+  "gates": "PASS: tested quality/correctness and both measured before/after rates improve; CIRU whole-request latency below Halo at both measured input lengths. See separate generation-rate tradeoffs.",
+  "quality": {
+    "Previous CIRU": {
+      "label": "v2-final-before",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    },
+    "CIRU v3": {
+      "label": "v3-final-C1",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    },
+    "Halo": {
+      "label": "halo-selected-A1-mtp3",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    }
+  },
+  "speed": {
+    "Previous CIRU": {
+      "4096": {
+        "rows": 2,
+        "PP": 392.00366125294556,
+        "TG": 22.51733524535651,
+        "TG_range": [
+          22.05163555418191,
+          23.003129150072414
+        ],
+        "TTFP_s": 10.69582200050354,
+        "request_s": 16.336945295333862,
+        "drafted": 588,
+        "accepted": 156,
+        "idle_ram_GiB": 100.72896957397461,
+        "peak_ram_GiB": 103.4262580871582,
+        "peak_GTT_GiB": 88.57872772216797,
+        "peak_VRAM_GiB": 0.5218048095703125,
+        "max_delta_ram_GiB": 2.1545333862304688
+      },
+      "65536": {
+        "rows": 1,
+        "PP": 284.4915510098569,
+        "TG": 13.332401115050645,
+        "TG_range": [
+          13.332401115050644,
+          13.332401115050644
+        ],
+        "TTFP_s": 230.45938777923584,
+        "request_s": 239.98778319358826,
+        "drafted": 283,
+        "accepted": 79,
+        "idle_ram_GiB": 100.59429168701172,
+        "peak_ram_GiB": 101.31758499145508,
+        "peak_GTT_GiB": 88.57872772216797,
+        "peak_VRAM_GiB": 0.5173988342285156,
+        "max_delta_ram_GiB": 0.7232933044433594
+      }
+    },
+    "CIRU v3": {
+      "4096": {
+        "rows": 6,
+        "PP": 455.6541833933926,
+        "TG": 24.599275541650545,
+        "TG_range": [
+          23.714691718344156,
+          25.19420366676837
+        ],
+        "TTFP_s": 9.245681206385294,
+        "request_s": 14.40951939423879,
+        "drafted": 1680,
+        "accepted": 480,
+        "idle_ram_GiB": 103.05811818440755,
+        "peak_ram_GiB": 106.01566696166992,
+        "peak_GTT_GiB": 90.2929458618164,
+        "peak_VRAM_GiB": 0.5292892456054688,
+        "max_delta_ram_GiB": 2.4259910583496094
+      },
+      "65536": {
+        "rows": 3,
+        "PP": 369.80548798995795,
+        "TG": 24.219011568105987,
+        "TG_range": [
+          24.166916898916444,
+          24.30324323910564
+        ],
+        "TTFP_s": 177.32060686747232,
+        "request_s": 182.56692719459534,
+        "drafted": 717,
+        "accepted": 258,
+        "idle_ram_GiB": 102.90898895263672,
+        "peak_ram_GiB": 103.66057205200195,
+        "peak_GTT_GiB": 90.2929458618164,
+        "peak_VRAM_GiB": 0.5248832702636719,
+        "max_delta_ram_GiB": 0.7370414733886719
+      }
+    },
+    "Halo": {
+      "4096": {
+        "rows": 4,
+        "PP": 381.49292683846284,
+        "TG": 35.30120650212089,
+        "TG_range": [
+          34.77293274914806,
+          35.57767783656343
+        ],
+        "TTFP_s": 11.094179451465607,
+        "request_s": 14.69240266084671,
+        "drafted": 528,
+        "accepted": 328,
+        "idle_ram_GiB": 102.4269323348999,
+        "peak_ram_GiB": 104.8825912475586,
+        "peak_GTT_GiB": 93.07597351074219,
+        "peak_VRAM_GiB": 1.85498046875,
+        "max_delta_ram_GiB": 2.7664871215820312
+      },
+      "65536": {
+        "rows": 2,
+        "PP": 263.42074609163245,
+        "TG": 23.278333535262185,
+        "TG_range": [
+          23.22488713070601,
+          23.33202649489396
+        ],
+        "TTFP_s": 248.91091644763947,
+        "request_s": 254.3688863515854,
+        "drafted": 276,
+        "accepted": 160,
+        "idle_ram_GiB": 103.14214324951172,
+        "peak_ram_GiB": 103.72273635864258,
+        "peak_GTT_GiB": 93.17262649536133,
+        "peak_VRAM_GiB": 1.9886512756347656,
+        "max_delta_ram_GiB": 0.597412109375
+      }
+    }
+  },
+  "change_vs": {
+    "Previous CIRU": {
+      "4096": {
+        "PP_percent": 16.23722644247847,
+        "TG_percent": 9.245944396210781,
+        "request_time_reduction_percent": 11.797957734764408
+      },
+      "65536": {
+        "PP_percent": 29.988214650756074,
+        "TG_percent": 81.65528743930224,
+        "request_time_reduction_percent": 23.92657460928912
+      }
+    },
+    "Halo": {
+      "4096": {
+        "PP_percent": 19.43974614929942,
+        "TG_percent": -30.31604871586294,
+        "request_time_reduction_percent": 1.925371044735702
+      },
+      "65536": {
+        "PP_percent": 40.385863101807075,
+        "TG_percent": 4.041002468749988,
+        "request_time_reduction_percent": 28.2274928301358
+      }
+    }
+  },
+  "full_capacity": {
+    "rows": 1,
+    "PP": 257.4360682269633,
+    "TG": 17.99524727099241,
+    "TG_range": [
+      17.99524727099241,
+      17.99524727099241
+    ],
+    "TTFP_s": 1017.3788130283356,
+    "request_s": 1024.442313671112,
+    "drafted": 214,
+    "accepted": 90,
+    "idle_ram_GiB": 102.63951873779297,
+    "peak_ram_GiB": 104.53916549682617,
+    "peak_GTT_GiB": 90.98435974121094,
+    "peak_VRAM_GiB": 0.5237464904785156,
+    "max_delta_ram_GiB": 1.8996467590332031
+  },
+  "platform": "NixOS x86-64 / ROCm10 / gfx1151; prior Ubuntu qualification is historical",
+  "weights_changed": false,
+  "ready_package_included": true
+}

benchmarks/v3.0.0/binary-identity.json

@@ -0,0 +1,392 @@
+{
+  "version": "3.0.0",
+  "platform": "Tested NixOS x86-64 / ROCm10 gfx1151 only",
+  "source_identity": {
+    "base_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+    "patch_sha256": "47711a875f7b5a56395ca4825f8ebfa80a36c0f2987bfe02e4f4f531a1d8cbf6",
+    "files": {
+      "ggml/src/ggml-cuda/fattn-tile.cuh": "89afad4509e4829281156795a8bc6252f2988b8def7b23726ff2bf11df2f5e3a",
+      "ggml/src/ggml-cuda/fattn.cu": "f65d5aa062af32cf623ba8831cb23ac97997a4813931bd4253b327dd292554a9",
+      "ggml/src/ggml-cuda/getrows.cu": "45eeb1beceeb8a80b1b8db42be413b6a7564b87177ab1d8828c9cb498e3e7bff",
+      "ggml/src/ggml-cuda/ggml-cuda.cu": "500f6c9fbb25558bde3a1705e39d801c46cc0feb343185672fb8ec74c6919247",
+      "ggml/src/ggml-cuda/mmq.cuh": "b45c2902c4e9170f85e058493f928719eed7f8b50348e652988ba05b59a86c6f",
+      "ggml/src/ggml-cuda/top-k.cu": "3362a25540fdb304e9ba91707a6dee874665ea4d8da61d74820a5284c31f39d8",
+      "ggml/src/ggml-cuda/top-k.cuh": "45566445c41b406a3f1b6e4c37b05c9e9dcc770f3b23b65ccb83ba6575316996",
+      "src/llama-graph.cpp": "c20fb02a4b29e2c06967f57126236d2d5532e59cc318e1f4629c4c83d713f176",
+      "src/llama-kv-cache.cpp": "a789d3e2a0d062fbe43f292da0fc27e3a1134ee5e4a199d44a03d2f0f6c88643",
+      "src/llama-kv-cells.h": "8a3f84c8ed9b7b9a004c6eb803ac6bf7d1ce104c84bf9104f399bdb059bca2e9",
+      "src/llama-memory-hybrid-idx.cpp": "19ea11aae77a927c1d2eebfb6357e6f01e562976077f9dafe5a0df6a56df0d44",
+      "src/llama-memory-hybrid-idx.h": "209f04181f581936070575c1195dedab607bc519c5f897022e8f32de368a3ae0",
+      "src/models/qwen4exp.cpp": "397b0748df4a622f5082cc6e6f4813bcf65e55d249f071e328a6bd37cb682069",
+      "tests/test-backend-ops.cpp": "a74be2d6c593734beaeadec04052794adb73bca3093dae42798dc318efd92883",
+      "tests/test-qsa-sequence.cpp": "8d14609998dc731dfb3f28ad07c2360c4e81213dfd9abc34b10f06e1e6d30a24"
+    },
+    "profile_files": {
+      "profiles/strix-halo-production.env": "7051ecb38492291dc93b0564807c6c60c837dae2a444acdb92b8fd39bbf822aa",
+      "scripts/ciru/run-server.sh": "d29677489306037a21d2d5ab06db5eacdf2883b084529a3cc2dced0a10094864"
+    },
+    "scope": "Compiled inference sources and focused tests; release docs and archive identity are recorded separately.",
+    "excluded_hybrid_components": [
+      "CIRU_KV_DEVICE",
+      "CIRU_QSA_HISTORY_OWNER",
+      "CIRU_QSA_PREFILL_OWNER",
+      "CIRU_QSA_PEER_KV",
+      "hot/cold expert ownership",
+      "scheduler overlap",
+      "CIRU_QSA_POOL_CACHE_PREFILL",
+      "CIRU_QSA_COMPACT_MASK"
+    ]
+  },
+  "selection": {
+    "time": 1788918535.353631,
+    "selected_binaries": "v3-pre-q5/bin restored to runtime-build/bin",
+    "reference_protocol": "runs/v3-mtp-screen-2/protocol.lock.json",
+    "source_patch_sha256": "47711a875f7b5a56395ca4825f8ebfa80a36c0f2987bfe02e4f4f531a1d8cbf6",
+    "rejected_candidate": "q5-three-column",
+    "verification": "Every inference binary/library hash equals the pre-candidate reference; all source/profile hashes match final-source-identity.json."
+  },
+  "inference_protocol": "v3-default6-confirm",
+  "files": [
+    {
+      "name": "libggml-base.so",
+      "bytes": 926848,
+      "symlink": "libggml-base.so.0"
+    },
+    {
+      "name": "libggml-base.so.0",
+      "bytes": 926848,
+      "symlink": "libggml-base.so.0.22.0"
+    },
+    {
+      "name": "libggml-base.so.0.22.0",
+      "bytes": 926848,
+      "sha256": "f88012f7cca9127c7119081f842d34f37e34f813b3fee5ae9156a80bf8243659",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libgomp.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libpthread.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libggml-cpu.so",
+      "bytes": 1388112,
+      "symlink": "libggml-cpu.so.0"
+    },
+    {
+      "name": "libggml-cpu.so.0",
+      "bytes": 1388112,
+      "symlink": "libggml-cpu.so.0.22.0"
+    },
+    {
+      "name": "libggml-cpu.so.0.22.0",
+      "bytes": 1388112,
+      "sha256": "a290300b1f253986fc2983765dc1f39630f4acd82a50ee28700c7ef756bbc84d",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgomp.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libpthread.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libggml-hip.so",
+      "bytes": 66914056,
+      "symlink": "libggml-hip.so.0"
+    },
+    {
+      "name": "libggml-hip.so.0",
+      "bytes": 66914056,
+      "symlink": "libggml-hip.so.0.22.0"
+    },
+    {
+      "name": "libggml-hip.so.0.22.0",
+      "bytes": 66914056,
+      "sha256": "4b9372fe8a1224163a79962e2647fee0e6d92131ae8de61288952bb07831ea32",
+      "elf_dynamic": [
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/toolchains/therock-gfx1151-10.0.0/lib:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib/rocm_sysdeps/lib:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib/llvm/lib:/nix/store/chqq8mpmpyfi9kgsngya71akv5xicn03-gcc-15.2.0-lib/lib:/nix/store/57iz36553175g3178pvxjij8z5rcsd4n-glibc-2.42-61/lib:/srv/llm/work/strix-v3-20260908/runtime-build/bin:]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libhipblas.so.3]",
+        "0x0000000000000001 (NEEDED)             Shared library: [librocblas.so.5]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libamdhip64.so.7]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libggml.so",
+      "bytes": 55192,
+      "symlink": "libggml.so.0"
+    },
+    {
+      "name": "libggml.so.0",
+      "bytes": 55192,
+      "symlink": "libggml.so.0.22.0"
+    },
+    {
+      "name": "libggml.so.0.22.0",
+      "bytes": 55192,
+      "sha256": "dabccabb93c1f22e9e3275eda95b5edb6613caddf314502763836b59cc764dd4",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libdl.so.2]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libllama-bench-impl.so",
+      "bytes": 454672,
+      "sha256": "03fac17aad52446e86f42e0fd099f397ac1b64ab2d8adb08e96b932985162050",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libllama-common.so",
+      "bytes": 5684704,
+      "symlink": "libllama-common.so.0"
+    },
+    {
+      "name": "libllama-common.so.0",
+      "bytes": 5684704,
+      "symlink": "libllama-common.so.0.3.0"
+    },
+    {
+      "name": "libllama-common.so.0.3.0",
+      "bytes": 5684704,
+      "sha256": "075cd447a5ae081a3509be554e8b98e9cd060655bb99c4e7726a916008efaa08",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [ld-linux-x86-64.so.2]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libllama-server-impl.so",
+      "bytes": 6813560,
+      "sha256": "35943dcee932959eb36d0c12d30cf69b70199fbb1bd40a1230032786fc585af6",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libmtmd.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [ld-linux-x86-64.so.2]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libllama.so",
+      "bytes": 4899352,
+      "symlink": "libllama.so.0"
+    },
+    {
+      "name": "libllama.so.0",
+      "bytes": 4899352,
+      "symlink": "libllama.so.0.3.0"
+    },
+    {
+      "name": "libllama.so.0.3.0",
+      "bytes": 4899352,
+      "sha256": "b2e8bd39d3d989020508e6798ae0b92f22d825dac60e905f62ba15c5fc6bc42e",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [ld-linux-x86-64.so.2]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "libmtmd.so",
+      "bytes": 1886576,
+      "symlink": "libmtmd.so.0"
+    },
+    {
+      "name": "libmtmd.so.0",
+      "bytes": 1886576,
+      "symlink": "libmtmd.so.0.3.0"
+    },
+    {
+      "name": "libmtmd.so.0.3.0",
+      "bytes": 1886576,
+      "sha256": "7a40775c4a5de8c85b03d30a80c4131157dff5fc9207261e16325042c085156c",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [ld-linux-x86-64.so.2]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": []
+    },
+    {
+      "name": "llama-bench",
+      "bytes": 15944,
+      "sha256": "d6c364d33d83b38d05838bcdee728479bc921ed27321816deb22d5f516261850",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-bench-impl.so]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": [
+        "[Requesting program interpreter: /nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib/ld-linux-x86-64.so.2]"
+      ]
+    },
+    {
+      "name": "llama-server",
+      "bytes": 15944,
+      "sha256": "71262d4bddd5699b89cef369233ddedd0097e0585d0a4610fca081f4595490e5",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-server-impl.so]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libmtmd.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": [
+        "[Requesting program interpreter: /nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib/ld-linux-x86-64.so.2]"
+      ]
+    },
+    {
+      "name": "test-backend-ops",
+      "bytes": 1308680,
+      "sha256": "b85fada96e62b160111d63b025af9b2f38676bfb8bc8ed57af7f91d16fdddd00",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": [
+        "[Requesting program interpreter: /nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib/ld-linux-x86-64.so.2]"
+      ]
+    },
+    {
+      "name": "test-batch-alloc",
+      "bytes": 350624,
+      "sha256": "200c567fc9e4ae0dc7952f2dabaec21ea203d55ca5b3efacc872b2b27f64a63a",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": [
+        "[Requesting program interpreter: /nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib/ld-linux-x86-64.so.2]"
+      ]
+    },
+    {
+      "name": "test-qsa-sequence",
+      "bytes": 70040,
+      "sha256": "94c9a5b2f9dcdbf38d87dc8683d5050df61734d55133ac364942838bbf192332",
+      "elf_dynamic": [
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama-common.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libllama.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-cpu.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-hip.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libggml-base.so.0]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]",
+        "0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]",
+        "0x000000000000001d (RUNPATH)            Library runpath: [/srv/llm/work/strix-v3-20260908/runtime-build/bin:/srv/llm/toolchains/therock-gfx1151-10.0.0/lib::/nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib:/nix/store/si4q3zks5mn5jhzzyri9hhd3cv789vlm-gcc-15.2.0-lib/lib]"
+      ],
+      "interpreter": [
+        "[Requesting program interpreter: /nix/store/fjkx1l5cnskzrqacf08z7i8z17256w0j-glibc-2.42-61/lib/ld-linux-x86-64.so.2]"
+      ]
+    }
+  ],
+  "time": 1788920951.604526,
+  "scope": "Server, bench, focused tests and shared libraries. Not a standalone portable Linux SDK; no CLI binary was built."
+}

benchmarks/v3.0.0/comparison.json

@@ -0,0 +1,371 @@
+{
+  "optional_mtp2": {
+    "4096": {
+      "rows": 4,
+      "PP": 453.08718324929157,
+      "TG": 29.386021312616883,
+      "TG_range": [
+        28.85317053177984,
+        29.90677798267714
+      ],
+      "TTFP_s": 9.292197465896606,
+      "request_s": 13.615030109882355,
+      "drafted": 448,
+      "accepted": 284,
+      "idle_ram_GiB": 102.58193302154541,
+      "peak_ram_GiB": 105.33094024658203,
+      "peak_GTT_GiB": 89.85221862792969,
+      "peak_VRAM_GiB": 0.5273895263671875,
+      "max_delta_ram_GiB": 2.2389984130859375
+    },
+    "65536": {
+      "rows": 2,
+      "PP": 373.0762034331891,
+      "TG": 24.880485510349693,
+      "TG_range": [
+        24.87814118364711,
+        24.882830278916934
+      ],
+      "TTFP_s": 175.76338636875153,
+      "request_s": 180.8699688911438,
+      "drafted": 232,
+      "accepted": 136,
+      "idle_ram_GiB": 102.45176124572754,
+      "peak_ram_GiB": 103.16925811767578,
+      "peak_GTT_GiB": 89.85221862792969,
+      "peak_VRAM_GiB": 0.5229873657226562,
+      "max_delta_ram_GiB": 0.7927474975585938
+    }
+  },
+  "coding_workload_rates": {
+    "Previous CIRU": {
+      "rows": 20,
+      "PP": 148.5313314164119,
+      "TG": 53.327449701233775,
+      "TG_range": [
+        47.918969023381464,
+        57.79477805890751
+      ],
+      "TTFP_s": 0.811668586730957,
+      "request_s": 3.774738335609436,
+      "drafted": 3162,
+      "accepted": 2664,
+      "idle_ram_GiB": 105.73584384918213,
+      "peak_ram_GiB": 108.11417388916016,
+      "peak_GTT_GiB": 88.58134460449219,
+      "peak_VRAM_GiB": 0.5263214111328125,
+      "max_delta_ram_GiB": 0.6338348388671875,
+      "generated_tokens": 3179,
+      "prompt_seconds": 15.181982,
+      "generation_seconds": 59.237784999999995,
+      "total_request_seconds": 75.49476671218872
+    },
+    "CIRU v3 MTP6": {
+      "rows": 20,
+      "PP": 219.50646380829988,
+      "TG": 53.2389694088275,
+      "TG_range": [
+        48.25151585152767,
+        57.44670920116761
+      ],
+      "TTFP_s": 0.5656813144683838,
+      "request_s": 3.533544683456421,
+      "drafted": 3162,
+      "accepted": 2664,
+      "idle_ram_GiB": 107.95920333862304,
+      "peak_ram_GiB": 110.46823501586914,
+      "peak_GTT_GiB": 90.29296112060547,
+      "peak_VRAM_GiB": 0.5336418151855469,
+      "max_delta_ram_GiB": 0.4556770324707031,
+      "generated_tokens": 3179,
+      "prompt_seconds": 10.273046,
+      "generation_seconds": 59.336235,
+      "total_request_seconds": 70.67089366912842
+    },
+    "CIRU v3 MTP2": {
+      "rows": 20,
+      "PP": 226.2516078412486,
+      "TG": 39.63335865299028,
+      "TG_range": [
+        36.41629100986472,
+        41.0896075498267
+      ],
+      "TTFP_s": 0.5504455924034118,
+      "request_s": 4.578490233421325,
+      "drafted": 2180,
+      "accepted": 2119,
+      "idle_ram_GiB": 107.43915119171143,
+      "peak_ram_GiB": 109.95774841308594,
+      "peak_GTT_GiB": 89.85222625732422,
+      "peak_VRAM_GiB": 0.530792236328125,
+      "max_delta_ram_GiB": 0.4302635192871094,
+      "generated_tokens": 3212,
+      "prompt_seconds": 9.96678,
+      "generation_seconds": 80.538216,
+      "total_request_seconds": 91.56980466842651
+    },
+    "Halo MTP3": {
+      "rows": 20,
+      "PP": 176.1027271635372,
+      "TG": 49.48345935868638,
+      "TG_range": [
+        46.09872313796549,
+        51.46158955201079
+      ],
+      "TTFP_s": 0.701993727684021,
+      "request_s": 3.9575579285621645,
+      "drafted": 2502,
+      "accepted": 2415,
+      "idle_ram_GiB": 106.45168437957764,
+      "peak_ram_GiB": 109.05359268188477,
+      "peak_GTT_GiB": 91.11787796020508,
+      "peak_VRAM_GiB": 1.8333740234375,
+      "max_delta_ram_GiB": 0.5126419067382812,
+      "generated_tokens": 3241,
+      "prompt_seconds": 12.805026,
+      "generation_seconds": 65.092458,
+      "total_request_seconds": 79.15115857124329
+    }
+  },
+  "speed": {
+    "Previous CIRU": {
+      "4096": {
+        "rows": 2,
+        "PP": 392.00366125294556,
+        "TG": 22.51733524535651,
+        "TG_range": [
+          22.05163555418191,
+          23.003129150072414
+        ],
+        "TTFP_s": 10.69582200050354,
+        "request_s": 16.336945295333862,
+        "drafted": 588,
+        "accepted": 156,
+        "idle_ram_GiB": 100.72896957397461,
+        "peak_ram_GiB": 103.4262580871582,
+        "peak_GTT_GiB": 88.57872772216797,
+        "peak_VRAM_GiB": 0.5218048095703125,
+        "max_delta_ram_GiB": 2.1545333862304688
+      },
+      "65536": {
+        "rows": 1,
+        "PP": 284.4915510098569,
+        "TG": 13.332401115050645,
+        "TG_range": [
+          13.332401115050644,
+          13.332401115050644
+        ],
+        "TTFP_s": 230.45938777923584,
+        "request_s": 239.98778319358826,
+        "drafted": 283,
+        "accepted": 79,
+        "idle_ram_GiB": 100.59429168701172,
+        "peak_ram_GiB": 101.31758499145508,
+        "peak_GTT_GiB": 88.57872772216797,
+        "peak_VRAM_GiB": 0.5173988342285156,
+        "max_delta_ram_GiB": 0.7232933044433594
+      }
+    },
+    "CIRU v3": {
+      "4096": {
+        "rows": 6,
+        "PP": 455.6541833933926,
+        "TG": 24.599275541650545,
+        "TG_range": [
+          23.714691718344156,
+          25.19420366676837
+        ],
+        "TTFP_s": 9.245681206385294,
+        "request_s": 14.40951939423879,
+        "drafted": 1680,
+        "accepted": 480,
+        "idle_ram_GiB": 103.05811818440755,
+        "peak_ram_GiB": 106.01566696166992,
+        "peak_GTT_GiB": 90.2929458618164,
+        "peak_VRAM_GiB": 0.5292892456054688,
+        "max_delta_ram_GiB": 2.4259910583496094
+      },
+      "65536": {
+        "rows": 3,
+        "PP": 369.80548798995795,
+        "TG": 24.219011568105987,
+        "TG_range": [
+          24.166916898916444,
+          24.30324323910564
+        ],
+        "TTFP_s": 177.32060686747232,
+        "request_s": 182.56692719459534,
+        "drafted": 717,
+        "accepted": 258,
+        "idle_ram_GiB": 102.90898895263672,
+        "peak_ram_GiB": 103.66057205200195,
+        "peak_GTT_GiB": 90.2929458618164,
+        "peak_VRAM_GiB": 0.5248832702636719,
+        "max_delta_ram_GiB": 0.7370414733886719
+      }
+    },
+    "Halo": {
+      "4096": {
+        "rows": 4,
+        "PP": 381.49292683846284,
+        "TG": 35.30120650212089,
+        "TG_range": [
+          34.77293274914806,
+          35.57767783656343
+        ],
+        "TTFP_s": 11.094179451465607,
+        "request_s": 14.69240266084671,
+        "drafted": 528,
+        "accepted": 328,
+        "idle_ram_GiB": 102.4269323348999,
+        "peak_ram_GiB": 104.8825912475586,
+        "peak_GTT_GiB": 93.07597351074219,
+        "peak_VRAM_GiB": 1.85498046875,
+        "max_delta_ram_GiB": 2.7664871215820312
+      },
+      "65536": {
+        "rows": 2,
+        "PP": 263.42074609163245,
+        "TG": 23.278333535262185,
+        "TG_range": [
+          23.22488713070601,
+          23.33202649489396
+        ],
+        "TTFP_s": 248.91091644763947,
+        "request_s": 254.3688863515854,
+        "drafted": 276,
+        "accepted": 160,
+        "idle_ram_GiB": 103.14214324951172,
+        "peak_ram_GiB": 103.72273635864258,
+        "peak_GTT_GiB": 93.17262649536133,
+        "peak_VRAM_GiB": 1.9886512756347656,
+        "max_delta_ram_GiB": 0.597412109375
+      }
+    }
+  },
+  "quality": {
+    "Previous CIRU": {
+      "label": "v2-final-before",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    },
+    "CIRU v3": {
+      "label": "v3-final-C1",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    },
+    "Halo": {
+      "label": "halo-selected-A1-mtp3",
+      "tasks": 20,
+      "base_pass": 20,
+      "plus_pass": 20,
+      "failed_tasks": [],
+      "recall": {
+        "8192": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 7947
+        },
+        "65536": {
+          "keys_present": true,
+          "cached_keys_present": true,
+          "exact_replay": true,
+          "cached_tokens": 65291
+        }
+      }
+    }
+  },
+  "change_vs": {
+    "Previous CIRU": {
+      "4096": {
+        "PP_percent": 16.23722644247847,
+        "TG_percent": 9.245944396210781,
+        "request_time_reduction_percent": 11.797957734764408
+      },
+      "65536": {
+        "PP_percent": 29.988214650756074,
+        "TG_percent": 81.65528743930224,
+        "request_time_reduction_percent": 23.92657460928912
+      }
+    },
+    "Halo": {
+      "4096": {
+        "PP_percent": 19.43974614929942,
+        "TG_percent": -30.31604871586294,
+        "request_time_reduction_percent": 1.925371044735702
+      },
+      "65536": {
+        "PP_percent": 40.385863101807075,
+        "TG_percent": 4.041002468749988,
+        "request_time_reduction_percent": 28.2274928301358
+      }
+    }
+  },
+  "full_capacity": {
+    "rows": 1,
+    "PP": 257.4360682269633,
+    "TG": 17.99524727099241,
+    "TG_range": [
+      17.99524727099241,
+      17.99524727099241
+    ],
+    "TTFP_s": 1017.3788130283356,
+    "request_s": 1024.442313671112,
+    "drafted": 214,
+    "accepted": 90,
+    "idle_ram_GiB": 102.63951873779297,
+    "peak_ram_GiB": 104.53916549682617,
+    "peak_GTT_GiB": 90.98435974121094,
+    "peak_VRAM_GiB": 0.5237464904785156,
+    "max_delta_ram_GiB": 1.8996467590332031
+  },
+  "scope": "local-custom; nonthinking; identical speed token input and published sampler; each package own profile; 262144 capacity; external speed input4096/65536; CIRU-only261888 full-capacity check",
+  "positions": {
+    "Previous CIRU": [
+      "v2-final-before"
+    ],
+    "CIRU v3": [
+      "v3-final-C1",
+      "v3-final-C2",
+      "v3-default6-confirm"
+    ],
+    "Halo": [
+      "halo-selected-A1-mtp3",
+      "halo-selected-A2-mtp3"
+    ]
+  },
+  "release_gate": "PASS: tested quality/correctness and both measured before/after rates improve; CIRU whole-request latency below Halo at both measured input lengths. See separate generation-rate tradeoffs."
+}

benchmarks/v3.0.0/default-mtp-decision.json

@@ -0,0 +1,22 @@
+{
+  "decision": "Retain MTP6 general default; MTP2 qualified as an optional long-request setting",
+  "trigger": "The high-acceptance short coding panel exposed a regression from changing the general default to MTP2.",
+  "coding_panel": {
+    "tasks": 20,
+    "previous_CIRU_MTP6_TG": 53.327449701233775,
+    "v3_MTP6_TG": 53.2389694088275,
+    "v3_MTP2_TG": 39.63335865299028,
+    "Halo_MTP3_TG": 49.48345935868638,
+    "previous_vs_v3_MTP6_token_streams_identical": 20,
+    "generated_tokens": {
+      "previous": 3179,
+      "v3_MTP6": 3179,
+      "v3_MTP2": 3212,
+      "Halo_MTP3": 3241
+    }
+  },
+  "quality": "All four runs20/20base+plus; same request panel, different sampled output lengths at different depths.",
+  "scope": "Default selection across two workload families; no adaptive algorithm implemented and no claim this is Halo optimal on every coding task.",
+  "queue_transition": "MTP2C2 is allowed to finish. The prior queue source/profile guard then deliberately holds subsequent inference because launcher default changed. New default-final unit verifies restored source/binaries and runs freshMTP6speed,HaloMTP3repeat,fullcapacityMTP6.",
+  "source_patch_unchanged": "47711a875f7b5a56395ca4825f8ebfa80a36c0f2987bfe02e4f4f531a1d8cbf6"
+}

benchmarks/v3.0.0/derived-cache-budget.json

@@ -0,0 +1,80 @@
+{
+  "indexer_K_norm_tensors": [
+    {
+      "name": "blk.3.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.7.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.11.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.15.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.19.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.23.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.27.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.31.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.35.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.39.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.43.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    },
+    {
+      "name": "blk.47.indexer.k_norm.weight",
+      "shape": [
+        128
+      ]
+    }
+  ],
+  "full_capacity": 262144,
+  "blocks": 65536,
+  "F32_minus_F16_bytes": 201326592,
+  "MiB": 192.0
+}

benchmarks/v3.0.0/git-source.json

@@ -0,0 +1,12 @@
+{
+  "version": "3.0.0",
+  "tag": "v3.0.0",
+  "commit": "e2b584751f89e02334304365edfe4c5404fd9e5c",
+  "tree": "106fbbf40b85acf8ad311ac0a5fdf801df3c23f3",
+  "base_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "archive": "ciru-runtime-v3.0.0-source.tar.gz",
+  "archive_sha256": "5680641c5f38eb5e8fe23f9af73bc4347b76c670e2425e1f6a2bbb9ebef1723a",
+  "file_count": 3615,
+  "manifest": "source-manifest.json",
+  "verification": "Every file content, executable mode and symlink target matches the release Git tree; named archive bytes are verified independently."
+}

benchmarks/v3.0.0/kernel-path-summary.json

@@ -0,0 +1,134 @@
+{
+  "all_dispatches": 159583,
+  "execution_paths": {
+    "ciru_top_k_radix_cells": {
+      "calls": 204,
+      "ms": 123.961048,
+      "tail_calls": 120
+    },
+    "ciru_get_rows_tiny_f32": {
+      "calls": 132,
+      "ms": 0.315596,
+      "tail_calls": 120
+    },
+    "flash_attn_tile<256, 256, 1, 4": {
+      "calls": 132,
+      "ms": 36.498432,
+      "tail_calls": 120
+    },
+    "set_rows": {
+      "calls": 1847,
+      "ms": 19.569685,
+      "tail_calls": 674
+    },
+    "rope": {
+      "calls": 1534,
+      "ms": 44.910432,
+      "tail_calls": 526
+    }
+  },
+  "tail_window_ms": 911.46,
+  "tail_scope": "Approximate decode tail: final kernel end minus native generation duration. Includes possible release work; not exact per-phase attribution.",
+  "tail_top": [
+    {
+      "kernel": "void mul_mat_vec_q_moe<(ggml_type)3, 2>(void const*, void const*, int const*, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 1296,
+      "ms": 139.791175
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)13, 3, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 2616,
+      "ms": 127.111853
+    },
+    {
+      "kernel": "void unary_gated_op_kernel<&(op_sigmoid(float)), float>(float const*, float const*, float*, long, long, long, long)",
+      "calls": 1532,
+      "ms": 72.728934
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)8, 3, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 2352,
+      "ms": 48.740039
+    },
+    {
+      "kernel": "void flash_attn_tile<256, 256, 1, 4, false, true, false>(char const*, char const*, char const*, char const*, char const*, int const*, float*, HIP_vector_type<float, 2u>*, float, float, float, float, unsigned int, float, int, HIP_vector_type<unsigned int, 3u>, int, int, int, int, int, int, int, int, int, int, int, long, int, int, long, int, int, int, int, int, long)",
+      "calls": 120,
+      "ms": 33.281598
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)8, 1, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 564,
+      "ms": 24.043484
+    },
+    {
+      "kernel": "ciru_top_k_radix_cells(float const*, int*, int, int)",
+      "calls": 120,
+      "ms": 17.603267
+    },
+    {
+      "kernel": "void mul_mat_vec_f<float, float, 3, 256, false, false>(float const*, float const*, int const*, ggml_cuda_mm_fusion_args_device, float*, int, HIP_vector_type<unsigned int, 3u>, int, int, int, HIP_vector_type<unsigned int, 3u>, int, int, int, HIP_vector_type<unsigned int, 3u>, int, int, int, int)",
+      "calls": 768,
+      "ms": 14.148268
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)13, 2, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 327,
+      "ms": 13.599575
+    },
+    {
+      "kernel": "quantize_q8_1(float const*, void*, long, long, long, long, long, unsigned int, HIP_vector_type<unsigned int, 3u>)",
+      "calls": 8368,
+      "ms": 12.83698
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)13, 1, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 231,
+      "ms": 11.117755
+    },
+    {
+      "kernel": "void gated_delta_net_cuda<128, false, true>(float const*, float const*, float const*, float const*, float const*, float const*, float*, float*, long, long, long, long, long, long, long, long, long, long, long, long, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, float, long, int)",
+      "calls": 360,
+      "ms": 10.458852
+    },
+    {
+      "kernel": "void k_bin_bcast<&(op_add(float, float)), float, float, float, float const*>(float const*, float const*, float*, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, float const*)",
+      "calls": 6104,
+      "ms": 9.455391
+    },
+    {
+      "kernel": "void concat_non_cont<unsigned int, 0>(char const*, char const*, char*, long, long, long, long, unsigned long, unsigned long, unsigned long, unsigned long, long, long, long, long, unsigned long, unsigned long, unsigned long, unsigned long, long, long, long, long, unsigned long, unsigned long, unsigned long, unsigned long)",
+      "calls": 333,
+      "ms": 9.070306
+    },
+    {
+      "kernel": "void k_get_rows_float_vec<float>(float const*, int const*, float*, long, long, HIP_vector_type<unsigned int, 3u>, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long)",
+      "calls": 360,
+      "ms": 6.998173
+    },
+    {
+      "kernel": "void k_bin_bcast<&(op_mul(float, float)), float, float, float, float const*>(float const*, float const*, float*, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, float const*)",
+      "calls": 3450,
+      "ms": 6.805671
+    },
+    {
+      "kernel": "__amd_rocclr_copyBufferRectAligned",
+      "calls": 2472,
+      "ms": 5.723215
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)8, 2, false, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 294,
+      "ms": 5.695476
+    },
+    {
+      "kernel": "void mul_mat_vec_q<(ggml_type)3, 1, true, false, false>(void const*, void const*, int const*, ggml_cuda_mm_fusion_args_device, float*, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, HIP_vector_type<unsigned int, 3u>, unsigned int, unsigned int, unsigned int, unsigned int)",
+      "calls": 48,
+      "ms": 4.986048
+    },
+    {
+      "kernel": "void k_argsort_f32_i32<(ggml_sort_order)1>(float const*, int*, int, int)",
+      "calls": 432,
+      "ms": 4.055824
+    }
+  ]
+}

benchmarks/v3.0.0/prior-package-archive.json

@@ -0,0 +1,6 @@
+{
+  "archive": "qsa-v2.0.1-prior-package.tar.gz",
+  "bytes": 80157489,
+  "sha256": "ae5b199b97981efddf8c649e183b169b2720e175d8a378d28d3a4299adfb3b06",
+  "scope": "Original READY.md plus exact release-ready and HF stages, including both checksum trees; historical qualification, not new v3 Ubuntu certification"
+}

benchmarks/v3.0.0/qsa-v2.0.1-prior-package.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae5b199b97981efddf8c649e183b169b2720e175d8a378d28d3a4299adfb3b06
+size 80157489

benchmarks/v3.0.0/ready-integration.json

@@ -0,0 +1,10 @@
+{
+  "status": "included",
+  "ready_source_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "all_QSA_changes_inherited": true,
+  "original_archives_verified": true,
+  "original_tag_not_pushed": "v2.0.1",
+  "delivery": "Combined v3 release; preserve v2.0 tags/weights/assets",
+  "prior_source_sha256": "28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5",
+  "prior_evidence_sha256": "a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66"
+}

benchmarks/v3.0.0/source-identity.json

@@ -0,0 +1,36 @@
+{
+  "base_commit": "9ea2390a71ae9f3d1cab519bbe099eb4ee06380e",
+  "patch_sha256": "47711a875f7b5a56395ca4825f8ebfa80a36c0f2987bfe02e4f4f531a1d8cbf6",
+  "files": {
+    "ggml/src/ggml-cuda/fattn-tile.cuh": "89afad4509e4829281156795a8bc6252f2988b8def7b23726ff2bf11df2f5e3a",
+    "ggml/src/ggml-cuda/fattn.cu": "f65d5aa062af32cf623ba8831cb23ac97997a4813931bd4253b327dd292554a9",
+    "ggml/src/ggml-cuda/getrows.cu": "45eeb1beceeb8a80b1b8db42be413b6a7564b87177ab1d8828c9cb498e3e7bff",
+    "ggml/src/ggml-cuda/ggml-cuda.cu": "500f6c9fbb25558bde3a1705e39d801c46cc0feb343185672fb8ec74c6919247",
+    "ggml/src/ggml-cuda/mmq.cuh": "b45c2902c4e9170f85e058493f928719eed7f8b50348e652988ba05b59a86c6f",
+    "ggml/src/ggml-cuda/top-k.cu": "3362a25540fdb304e9ba91707a6dee874665ea4d8da61d74820a5284c31f39d8",
+    "ggml/src/ggml-cuda/top-k.cuh": "45566445c41b406a3f1b6e4c37b05c9e9dcc770f3b23b65ccb83ba6575316996",
+    "src/llama-graph.cpp": "c20fb02a4b29e2c06967f57126236d2d5532e59cc318e1f4629c4c83d713f176",
+    "src/llama-kv-cache.cpp": "a789d3e2a0d062fbe43f292da0fc27e3a1134ee5e4a199d44a03d2f0f6c88643",
+    "src/llama-kv-cells.h": "8a3f84c8ed9b7b9a004c6eb803ac6bf7d1ce104c84bf9104f399bdb059bca2e9",
+    "src/llama-memory-hybrid-idx.cpp": "19ea11aae77a927c1d2eebfb6357e6f01e562976077f9dafe5a0df6a56df0d44",
+    "src/llama-memory-hybrid-idx.h": "209f04181f581936070575c1195dedab607bc519c5f897022e8f32de368a3ae0",
+    "src/models/qwen4exp.cpp": "397b0748df4a622f5082cc6e6f4813bcf65e55d249f071e328a6bd37cb682069",
+    "tests/test-backend-ops.cpp": "a74be2d6c593734beaeadec04052794adb73bca3093dae42798dc318efd92883",
+    "tests/test-qsa-sequence.cpp": "8d14609998dc731dfb3f28ad07c2360c4e81213dfd9abc34b10f06e1e6d30a24"
+  },
+  "profile_files": {
+    "profiles/strix-halo-production.env": "7051ecb38492291dc93b0564807c6c60c837dae2a444acdb92b8fd39bbf822aa",
+    "scripts/ciru/run-server.sh": "d29677489306037a21d2d5ab06db5eacdf2883b084529a3cc2dced0a10094864"
+  },
+  "scope": "Compiled inference sources and focused tests; release docs and archive identity are recorded separately.",
+  "excluded_hybrid_components": [
+    "CIRU_KV_DEVICE",
+    "CIRU_QSA_HISTORY_OWNER",
+    "CIRU_QSA_PREFILL_OWNER",
+    "CIRU_QSA_PEER_KV",
+    "hot/cold expert ownership",
+    "scheduler overlap",
+    "CIRU_QSA_POOL_CACHE_PREFILL",
+    "CIRU_QSA_COMPACT_MASK"
+  ]
+}

benchmarks/v3.0.0/strix-v3.0.0-evidence.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de1d57fa7dc841eb0573c3f3d2cf43958e78d4cafd88e568f9987a73eacd9e42
+size 12681911

profiles/strix-halo-production.env

@@ -9,3 +9,17 @@ GGML_QSA_LONG_TOPK=1
 GGML_QSA_RESTORE_FAST=1
 CIRU_MTP_TOPK10=1
 CIRU_MTP_SHORTLIST=32768
+
+# Version 3 portable Strix kernels and derived QSA history. Set a switch to 0 to disable it.
+GGML_CUDA_Q41_MOE_FORCE_J_ALL_M=${GGML_CUDA_Q41_MOE_FORCE_J_ALL_M:-1}
+GGML_QSA_PREFILL_WIDE=${GGML_QSA_PREFILL_WIDE:-1}
+GGML_QSA_ALL_ROWS=${GGML_QSA_ALL_ROWS:-1}
+CIRU_QSA_WARP_SCAN=${CIRU_QSA_WARP_SCAN:-1}
+CIRU_QSA_RADIX_SELECT=${CIRU_QSA_RADIX_SELECT:-1}
+CIRU_QSA_INDEXED_DECODE=${CIRU_QSA_INDEXED_DECODE:-1}
+CIRU_PLE_PREV_DIRECT=${CIRU_PLE_PREV_DIRECT:-1}
+CIRU_QSA_POOL_CACHE=${CIRU_QSA_POOL_CACHE:-1}
+CIRU_QSA_TINY_GATHER=${CIRU_QSA_TINY_GATHER:-1}
+
+# Experimental draft-only window; the qualified profile retains full draft context.
+CIRU_MTP_ATTENTION_WINDOW=${CIRU_MTP_ATTENTION_WINDOW:-0}

runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5
+size 36780030

runtime/v3.0.0/ciru-runtime-v3.0.0-nixos-gfx1151.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee18044bba8eb5e314bd89a1898eaa5a387a60c7d7feb7f92248c69c2dadeb50
+size 22668322

runtime/v3.0.0/ciru-runtime-v3.0.0-source.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5680641c5f38eb5e8fe23f9af73bc4347b76c670e2425e1f6a2bbb9ebef1723a
+size 36822527

v3.0.0-checksums.sha256

@@ -0,0 +1,35 @@
+c16fda8809ad10f0d25d2243f4f83e76c29b80e0ee1622a4357a6f7235df97b4  .gitattributes
+79cec9000309c18e78ebd55d4fe38f8cf01732634cebb43eb57b138939d4afd6  README.md
+3a46a7fd054305b2c1a931682015d93094312fb3b76c8ef83a2adb350b0f584c  assets/qwen38-flash-ciru-strix-iu4.jpg
+e260f72cba14e4c4e523794345f7de001444626240ab620d96dc44ee51beee90  assets/v2-bf16-fidelity.png
+88dd940fda0790ec6f474dc0a99019703114df253758ee14136c9686e93ea404  assets/v2-context-sweep.png
+2e1a7aa972d016ea70d3853d9328386717cb17f8802b02cbe9dae511484ca8a4  benchmarks/v2.0.1/FINAL-QUALIFICATION.json
+9d9e9d08ab443b616f2c693b08028f902cfee0fbb56a5ef997dd99183bf98fc0  benchmarks/v2.0.1/QUALIFICATION.md
+2f83bea64e1b302916bcc46f480c3b1040a3a5b5a9501cb73b11f63f32f96859  benchmarks/v2.0.1/git-source.json
+e3c728faab498ac46872c4281c7ba99b28ff7eedab95dc1da56de38c7cf602c6  benchmarks/v2.0.1/publication/INDEPENDENT-PACKAGE-REVIEW.json
+f4697dbabfde5778199595267189f4662eb637e7db20ad43bc95c3420ef15fe2  benchmarks/v2.0.1/publication/PACKAGE-VERIFICATION.json
+8ca9332e0779b66d407afd075d80c5876d94fda42e504095a800a786c063ff69  benchmarks/v2.0.1/publication/READY.md
+d73bafff54fdf9adb8754816d33d2f366811743b7f42a48f38db01a836e6fbd8  benchmarks/v2.0.1/publication/STAGED-CHECKSUMS.sha256
+c81afc287d03d0375409dbe0d614bb1931b344aa20769a360853c9582dff0159  benchmarks/v2.0.1/publication/STAGED-MODEL-CARD.md
+a032e4731db988b2b787563a4094ae662e73b3ad44488fda089c998998e1db66  benchmarks/v2.0.1/qsa-v2.0.1-evidence.tar.gz
+b93c6f1454f2a25d0fe346a8c2d1955f04a0d35c853cde6e4e446fbd3a955e5b  benchmarks/v2.0.1/source-manifest.json
+9d1c30d515572a02d578359be0a6f74998eb2afc44cfe41ab0771bc338054df6  benchmarks/v3.0.0/COMPARISON.md
+a98b0b861595e3c350a0ceb91af2cf87609724f568bdad12271153f1af0fa77c  benchmarks/v3.0.0/FINAL-QUALIFICATION.json
+76e8c4c56d98add1c3fe25da595687f97d807f714e0bb60c0bed05d091c3c6ae  benchmarks/v3.0.0/binary-identity.json
+4d2a01e1c72445373f34a3768afb22541555b79a02d4119eaf301093554e9d5e  benchmarks/v3.0.0/comparison.json
+b6e7763c45fb930c5d11c07980db0cb1606f818ff6c2de00d94eeeb0af090dd2  benchmarks/v3.0.0/default-mtp-decision.json
+be591a43fc2af93c9ee0e95cd1bc3ef16b05828e3979e84d11eddd30f3777b5c  benchmarks/v3.0.0/derived-cache-budget.json
+de86ae4f64770ab6bdd387e903202573850ee31880ada0bb77288e7ee17bb42e  benchmarks/v3.0.0/git-source.json
+3834b7c615511ca1bc966028ac82bca0540ebc32a861ad769f014a0909b6e346  benchmarks/v3.0.0/kernel-path-summary.json
+911757a1d82c179bf3259140baf4fe3341ecd78c7b445c59b0872d82e1ebd5d6  benchmarks/v3.0.0/prior-package-archive.json
+ae5b199b97981efddf8c649e183b169b2720e175d8a378d28d3a4299adfb3b06  benchmarks/v3.0.0/qsa-v2.0.1-prior-package.tar.gz
+619b3414541ac584b36e9fc3ccc6ba3ccde3fe9814fd6ca0e2937a6de90d81a1  benchmarks/v3.0.0/ready-integration.json
+e5f81ee2e39c71973c23b6297493b7a6391bd9eb72ed034d5e291ef2585f6da8  benchmarks/v3.0.0/source-identity.json
+2f29280aa186703dbb657bc34f2209055c107c8ba299a3889f9de1e4702597fe  benchmarks/v3.0.0/source-manifest.json
+de1d57fa7dc841eb0573c3f3d2cf43958e78d4cafd88e568f9987a73eacd9e42  benchmarks/v3.0.0/strix-v3.0.0-evidence.tar.gz
+cd1177aef73724230e04e53171321339177a7fcb6cd1292d6794cb9abfa98f54  checksums.sha256
+7051ecb38492291dc93b0564807c6c60c837dae2a444acdb92b8fd39bbf822aa  profiles/strix-halo-production.env
+28b2194323ba1105921f2c52639cc28cce8b8336f4f0a3a54349c249823fa7c5  runtime/v2.0.1/ciru-runtime-v2.0.1-source.tar.gz
+ee18044bba8eb5e314bd89a1898eaa5a387a60c7d7feb7f92248c69c2dadeb50  runtime/v3.0.0/ciru-runtime-v3.0.0-nixos-gfx1151.tar.gz
+5680641c5f38eb5e8fe23f9af73bc4347b76c670e2425e1f6a2bbb9ebef1723a  runtime/v3.0.0/ciru-runtime-v3.0.0-source.tar.gz
+73bc5f5c4c99eac1670ecd2ecafdd0138829e3d88a7868178a837157d7c8d98e  v2.0-checksums.sha256