d2/K256, 14.45 GiB: KL 40.3 vs the 4-bit conversion's 45.8, at 0.50 GiB smaller

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+qwen38_ladder.png filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,131 @@
+---
+language:
+- en
+license: apache-2.0
+library_name: mlx
+pipeline_tag: text-generation
+base_model: Qwen/Qwen3.8-27B
+base_model_relation: quantized
+tags:
+- mlx
+- quantized
+- vector-quantization
+- apple-silicon
+- qwen3.8
+---
+
+# TheDrainFlorist/Qwen3.8-27B-VQ-4.5bpw
+
+**14.5 GiB — smaller than our 4-bit conversion, and closer to bf16.**
+
+A vector-quantized build of
+[Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) for Apple Silicon.
+Stock `mlx-lm`, no patches — the VQ runtime ships inside the checkpoint as
+`model.py`.
+
+At the time of release no MLX-format quantization of this model had been
+published, so the affine builds compared against below are our own
+conversions rather than community artifacts. That is a weaker class of
+evidence — a comparator you build yourself can be built badly, and one of
+ours was; see Comparators.
+
+![where these releases sit](qwen38_ladder.png)
+
+## Measured results
+
+Scored against the bf16 teacher on the same corpus with an unmodified
+`mlx-lm`. All sizes include the 333-tensor bf16 vision tower (0.859 GiB),
+carried by every build here.
+
+| build | size | KL to bf16 (mnats/tok) | top-1 agreement | perplexity |
+|---|---|---|---|---|
+| affine q2 (ours) | 8.69 GiB | 1426.9 | 46.1% | 16.435 |
+| affine q3 (ours) | 11.82 GiB | 187.8 | 79.5% | 5.832 |
+| **this model** | **14.45 GiB** | **40.3** | **90.1%** | 5.233 |
+| affine q4 (ours) | 14.95 GiB | 45.8 | 89.8% | 5.206 |
+| affine q6 (ours) | 21.21 GiB | 3.71 | 96.8% | 5.260 |
+| affine q8 (ours) | 27.48 GiB | 1.25 | 98.5% | 5.241 |
+| bf16 | 51.7 GiB | 0 | 100% | — |
+
+This is the cleanest comparison on the ladder: against the 4-bit affine
+conversion it is **0.50 GiB smaller and 12% closer to bf16** (40.3 millinats
+against 45.8), with 0.3 points better token agreement. Smaller and better on
+the same instrument, no trade to weigh.
+
+**Rank these by KL, not perplexity.** On this instruction-tuned family
+perplexity barely moves — every build from 3-bit upward sits between 5.19 and
+5.35, inside the measurement's own noise — while divergence from the teacher
+moves by a factor of forty across the same range. Perplexity is an aggregate
+over finite text and absorbs offsetting errors; KL measures distance to the
+teacher's distribution directly.
+
+## Runtime
+
+**Not measured on this artifact.** No decode or prefill benchmark has been
+run on this build, and quoting a sibling's figures would be a substitution
+this project does not make. Resident memory is about 13.60 GiB — the disk
+figure less the vision tower, which `mlx-lm` does not load.
+
+Runs on a 16 GB, tightly sized machine.
+
+## Run it
+
+```bash
+pip install mlx-lm
+python -m mlx_lm generate \
+  --model TheDrainFlorist/Qwen3.8-27B-VQ-4.5bpw \
+  --prompt "Explain vector quantization briefly." \
+  --max-tokens 512
+```
+
+## How it was built
+
+Vector quantization of the dense MLP trio at **d=2, K=256**. Each 2-weight
+subvector stores one 8-bit index into a per-tensor 256-entry fp16 codebook. With an fp16 scale
+per (row, 64 weights) that comes to 4.25 bits per weight over the quantized
+surface; everything else in the model is 8-bit.
+
+Every quantized tensor uses this one geometry: no depth schedule, no mixed
+allocation.
+
+**Codebooks are fit in pure weight space** — k-means over the weight
+subvectors, no Hessian, no activation statistics, no calibration corpus.
+
+**The fit is not seeded.** k-means draws an unseeded subsample, so this
+artifact is reproducible in recipe and geometry but not bit-for-bit. Margins
+are therefore quoted against a measured fit-to-fit floor rather than against a
+repeated build; on this family that floor is 2.085 millinats.
+
+## Comparators
+
+The affine rungs above are local conversions made with `mlx_lm.convert`. One
+correction worth stating plainly, because it was ours: the 8-bit comparator
+originally used here was not a uniform 8-bit build at all — its configuration
+declared a 4-bit default with per-module overrides, including the output head
+at 6 bits. It was rebuilt with defaults and re-scored, and the figure above is
+the rebuilt one. The bar moved *against* us when corrected, from 1.64 to 1.25
+millinats.
+
+## Where this stops paying
+
+Above roughly 5 bits per weight the advantage reverses on this model: our
+6-bit affine conversion reaches 3.7 millinats at 21.2 GiB, which no VQ rung we
+measured approaches at that size. Builds larger than the ones released here
+were measured and deliberately not published for that reason.
+
+## Verification
+
+Every tensor was decoded **from the published artifact** and compared against
+the bf16 source; no tensor exceeds 3x the artifact's own median reconstruction
+error. The bundled runtime was exercised as the executing copy in a stock
+venv, not merely present in the folder. Vision tower grafted from the base
+checkpoint and verified key-for-key against the official index, including the
+channels-last patch-embedding layout that a naive rename gets silently wrong.
+
+## Limitations
+
+- Perplexity cannot rank builds on this family — see above.
+- No throughput measurement, and no task-suite scores, for this artifact.
+- The affine comparators are our own conversions, not community builds.
+- Above ~5 bpw affine wins outright on this model; this collection stops
+  below that line deliberately.

chat_template.jinja

@@ -0,0 +1,170 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- set reasoning_instructions = '' %}
+{%- if enable_thinking is undefined or enable_thinking is true %}
+    {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
+    {%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %}
+        {{- raise_exception('Unexpected reasoning effort ' ~ reasoning_effort ~ '. Supported types are xhigh (default), medium, and low.') }}
+    {%- endif %}
+    {%- if resolved_reasoning_effort == 'xhigh' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
+    {%- elif resolved_reasoning_effort == 'low' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}
+    {%- endif %}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {%- if reasoning_instructions %}
+        {{- reasoning_instructions + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '<|im_start|>system\n' + (reasoning_instructions + '\n\n' if reasoning_instructions else '')  + content + '<|im_end|>\n' }}
+        {%- elif reasoning_instructions %}
+            {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+        {%- endif %}
+    {%- elif reasoning_instructions %}
+        {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if preserve_thinking is undefined or preserve_thinking is true or loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined and tool_call.arguments != '' %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,1694 @@
+{
+ "architectures": [
+  "Qwen3_5ForConditionalGeneration"
+ ],
+ "eos_token_id": [
+  248046,
+  248044
+ ],
+ "image_token_id": 248056,
+ "language_model_only": false,
+ "model_type": "qwen3_5",
+ "quantization": {
+  "group_size": 64,
+  "bits": 4,
+  "mode": "affine"
+ },
+ "quantization_config": {
+  "group_size": 64,
+  "bits": 4,
+  "mode": "affine"
+ },
+ "text_config": {
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attn_output_gate": true,
+  "bos_token_id": 248044,
+  "dtype": "bfloat16",
+  "eos_token_id": 248044,
+  "full_attention_interval": 4,
+  "head_dim": 256,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 17408,
+  "layer_types": [
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention",
+   "linear_attention",
+   "linear_attention",
+   "linear_attention",
+   "full_attention"
+  ],
+  "linear_conv_kernel_dim": 4,
+  "linear_key_head_dim": 128,
+  "linear_num_key_heads": 16,
+  "linear_num_value_heads": 48,
+  "linear_value_head_dim": 128,
+  "mamba_ssm_dtype": "float32",
+  "max_position_embeddings": 262144,
+  "model_type": "qwen3_5_text",
+  "mtp_num_hidden_layers": 1,
+  "mtp_use_dedicated_embeddings": false,
+  "num_attention_heads": 24,
+  "num_hidden_layers": 64,
+  "num_key_value_heads": 4,
+  "output_gate_type": "swish",
+  "pad_token_id": null,
+  "partial_rotary_factor": 0.25,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+   "mrope_interleaved": true,
+   "mrope_section": [
+    11,
+    11,
+    10
+   ],
+   "partial_rotary_factor": 0.25,
+   "rope_theta": 10000000,
+   "type": "default"
+  },
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "vocab_size": 248320
+ },
+ "tie_word_embeddings": false,
+ "transformers_version": "5.8.0.dev0",
+ "video_token_id": 248057,
+ "vision_end_token_id": 248054,
+ "vision_start_token_id": 248053,
+ "vq_linear": {
+  "language_model.model.layers.0.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.0.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.0.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.1.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.1.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.1.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.2.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.2.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.2.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.3.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.3.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.3.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.4.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.4.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.4.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.5.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.5.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.5.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.6.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.6.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.6.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.7.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.7.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.7.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.8.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.8.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.8.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.9.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.9.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.9.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.10.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.10.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.10.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.11.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.11.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.11.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.12.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.12.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.12.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.13.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.13.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.13.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.14.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.14.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.14.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.15.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.15.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.15.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.16.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.16.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.16.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.17.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.17.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.17.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.18.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.18.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.18.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.19.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.19.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.19.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.20.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.20.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.20.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.21.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.21.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.21.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.22.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.22.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.22.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.23.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.23.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.23.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.24.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.24.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.24.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.25.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.25.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.25.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.26.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.26.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.26.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.27.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.27.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.27.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.28.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.28.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.28.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.29.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.29.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.29.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.30.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.30.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.30.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.31.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.31.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.31.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.32.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.32.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.32.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.33.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.33.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.33.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.34.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.34.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.34.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.35.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.35.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.35.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.36.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.36.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.36.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.37.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.37.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.37.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.38.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.38.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.38.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.39.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.39.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.39.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.40.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.40.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.40.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.41.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.41.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.41.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.42.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.42.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.42.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.43.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.43.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.43.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.44.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.44.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.44.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.45.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.45.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.45.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.46.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.46.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.46.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.47.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.47.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.47.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.48.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.48.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.48.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.49.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.49.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.49.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.50.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.50.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.50.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.51.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.51.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.51.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.52.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.52.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.52.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.53.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.53.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.53.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.54.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.54.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.54.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.55.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.55.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.55.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.56.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.56.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.56.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.57.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.57.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.57.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.58.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.58.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.58.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.59.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.59.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.59.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.60.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.60.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.60.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.61.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.61.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.61.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.62.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.62.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.62.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.63.mlp.gate_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.63.mlp.up_proj": {
+   "experts": 1,
+   "out": 17408,
+   "in": 5120,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  },
+  "language_model.model.layers.63.mlp.down_proj": {
+   "experts": 1,
+   "out": 5120,
+   "in": 17408,
+   "k": 256,
+   "dim": 2,
+   "group": 64
+  }
+ },
+ "vq_embed": {},
+ "model_file": "model.py",
+ "vision_config": {
+  "deepstack_visual_indexes": [],
+  "depth": 27,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "in_channels": 3,
+  "initializer_range": 0.02,
+  "intermediate_size": 4304,
+  "model_type": "qwen3_5",
+  "num_heads": 16,
+  "num_position_embeddings": 2304,
+  "out_hidden_size": 5120,
+  "patch_size": 16,
+  "spatial_merge_size": 2,
+  "temporal_patch_size": 2
+ }
+}

generation_config.json

@@ -0,0 +1,12 @@
+{
+    "bos_token_id": 248044,
+    "do_sample": true,
+    "eos_token_id": [
+        248046,
+        248044
+    ],
+    "pad_token_id": 248044,
+    "temperature": 1.0,
+    "top_k": 20,
+    "top_p": 0.95
+}

model-00001.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af46d76bc45897ddc660284fd8e2f66f3688c3e7d6cacaa2e40e8c1f2108985a
+size 2119658552

model-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:add254e6ecb136fd0e90a954623df22893165f54c56b8e8b2637676b284a84c6
+size 1594027965

model-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4364cae7b25d4d440256bcaa81004f12af6f2fd9192db2a99ab869b576f0454a
+size 1793355260

model-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c0a6c18ec3c82aa4720068db65919538fb17d5e45535b4e06486a8ae4d28874
+size 9091424014

model-vision-graft.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f71b3e3e66250f3990e0943fc42082622cc9dc28538a529d0da7ee98723d8f4
+size 921497299

model.py

@@ -0,0 +1,268 @@
+"""Dense VQ modules — VQLinear and VQEmbedding.
+
+WHY THESE EXIST. vq_switch.VQSwitchLinear is EXPERT-shaped: its __call__
+takes routing indices and its kernels index a leading expert axis. A dense
+model (gemma-4-e4b) has neither, so shipping a VQ e4b needs drop-ins for
+plain nn.Linear and nn.Embedding.
+
+THE EMBEDDING IS THE GOOD CASE. A [262144, 10752] table is 35.5% of e4b's
+bytes, and an embedding lookup only ever needs the ROWS a batch touches —
+so VQEmbedding gathers those code rows and decodes just those, never
+materialising the table. Memory saving with no throughput cost.
+
+THE LINEAR IS THE HONEST CASE. A matmul needs the whole weight, so
+VQLinear decodes W per call: codebook[codes] -> [OUT, NSUB, D] ->
+[OUT, IN], scaled group-wise. That trades compute for resident memory and
+will be SLOWER than an 8-bit matmul; measure before believing any speed
+claim. Correctness first — a fused dense kernel is a later optimisation,
+and per E62 the fused work has to reproduce the scored numbers exactly
+before it may replace this path.
+
+Both mirror the fitter's contract exactly: scales are fp16 max-abs per
+group of G along `in`, codes index a [K, D] fp16 codebook.
+"""
+import os
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+def _decode(codes, codebook, scales, group_size, out_d, in_d):
+    """codes [R, NSUB] -> dense [R, in_d] fp16, scaled group-wise."""
+    w = codebook[codes.reshape(-1)].reshape(out_d, in_d)
+    w = (w.reshape(out_d, in_d // group_size, group_size)
+         * scales[..., None]).reshape(out_d, in_d)
+    return w
+
+
+def _unpack_rows(packed, nsub, bits):
+    """mlx-native inverse of vq_pack.pack for a [R, WPR] uint32 slab.
+
+    Runs on-GPU at gather time so packed EMBEDDING rows never round-trip
+    through numpy. Same 32-codes-per-block layout as vq_pack.py; verified
+    against vq_pack.unpack bit-exactly before first use.
+    """
+    R, wpr = packed.shape
+    nblk = nsub // 32
+    src = packed.reshape(R, nblk, bits).astype(mx.uint64)
+    mask = mx.array(( 1 << bits) - 1, dtype=mx.uint64)
+    outs = []
+    for i in range(32):
+        off = i * bits
+        w, sh = divmod(off, 32)
+        v = src[:, :, w] >> mx.array(sh, dtype=mx.uint64)
+        if sh + bits > 32:
+            v = v | (src[:, :, w + 1] << mx.array(32 - sh, dtype=mx.uint64))
+        outs.append((v & mask).astype(mx.uint32))
+    # outs[i] is code i of every block -> interleave back to row order
+    return mx.stack(outs, axis=-1).reshape(R, nsub)
+
+
+class VQLinear(nn.Module):
+    """Drop-in for a bias-free nn.Linear whose weight is VQ-coded.
+
+    codes may be unpacked ([OUT, NSUB] uint8/uint16) or packed
+    ([OUT, WPR] uint32 of pack_bits-wide fields, vq_pack.py layout).
+    """
+
+    def __init__(self, codes, codebook, vq_scales, group_size=64,
+                 pack_bits=0, in_features=None):
+        super().__init__()
+        self.codes = codes
+        self.codebook = codebook
+        self.vq_scales = vq_scales
+        self.group_size = group_size
+        self.pack_bits = pack_bits
+        if pack_bits and in_features is None:
+            raise ValueError("packed codes need explicit in_features")
+        self._in_features = in_features
+        self._k_expect = int(codebook.shape[0])
+        self.freeze()
+
+    @property
+    def input_dims(self):
+        if self.pack_bits:
+            return self._in_features
+        return self.codes.shape[-1] * self.codebook.shape[1]
+
+    @property
+    def output_dims(self):
+        return self.codes.shape[0]
+
+    def __call__(self, x):
+        # Same sharding guard as VQSwitchLinear: a sliced codebook does not
+        # error, it silently decodes garbage (exo PR #2268).
+        if int(self.codebook.shape[0]) != self._k_expect:
+            raise RuntimeError(
+                f"VQ codebook was sharded: K={self.codebook.shape[0]}, "
+                f"expected {self._k_expect}. The codebook is a SHARED lookup "
+                f"table and must be REPLICATED across tensor-parallel ranks.")
+        OUT = self.codes.shape[0]
+        IN = (self._in_features if self.pack_bits
+              else self.codes.shape[1] * self.codebook.shape[1])
+        # A dense linear IS an expert layer with E=1 and every token routed
+        # to expert 0 — so the small-N path rides the PRODUCTION-VALIDATED
+        # fused kernels from vq_switch (E62: bit-identical, ~3x decode)
+        # instead of decoding the whole weight per call, which measured
+        # 11.5 tok/s vs the incumbent's 84 (2026-08-19). Prefill (large N)
+        # amortises one full decode across the whole batch, which is faster
+        # than the fused gather at that shape — same split vq_switch uses.
+        if self.pack_bits:
+            IN = self._in_features
+        orig_shape = x.shape
+        xf = x.reshape(-1, IN)
+        N = xf.shape[0]
+        # d=2 is the only geometry with a dense fused kernel. The expert
+        # kernel (_fused, E=1) was tried as the d!=2 small-N path and DIED AT
+        # KERNEL LOAD on the first real dense model: at 27B mlp shapes
+        # (IN 17408) its threadgroup allocation is 36864 bytes vs Metal's
+        # 32768 cap — shapes the 397B's experts never produce. Caught by
+        # III.10 smoke-gen (E95). Until a dense kernel exists for d>2,
+        # small-N takes the decode path: bit-exact, just slow at decode
+        # (~11.5 tok/s measured 08-19) — fine for gates and prefill-shaped
+        # scoring, NOT a shipping configuration.
+        # THREADGROUP CEILING (2026-08-21, E124). The d=2 dense kernel ALSO
+        # dies at kernel load on 27B mlp shapes: NSUB = IN/d = 8704 needs
+        # 35840 B of threadgroup memory against Metal's 32768 cap. So neither
+        # d=2 NOR d>2 has a usable fused dense path at this model's widths —
+        # the 397B's expert shapes simply never produce them. Measured, not
+        # inferred: the failure is a RuntimeError naming both numbers.
+        # Falling back to decode keeps the artifact CORRECT (bit-exact, the
+        # same path that produced its scores) at decode speed. This is not a
+        # performance fix and must not be read as one; a dense kernel that
+        # fits the cap is the real fix and does not exist yet.
+        _nsub = IN // int(self.codebook.shape[1])
+        _fits_tg = _nsub * 4 + 1024 <= 32768
+        if N <= 32 and int(self.codebook.shape[1]) == 2 and _fits_tg:
+            # DENSE fused kernel (2026-08-19): one simdgroup per output row
+            # instead of one thread, no expert axis. Bit-identical to the
+            # expert-kernel path below (the kernel replicates its float
+            # ordering exactly — verified, and the E62 KL number reproduces).
+            # SCOUT_VQ_DENSE_REF=1 keeps the old expert-shaped path callable
+            # as the reference for A/B checks.
+            if os.environ.get("SCOUT_VQ_DENSE_REF"):
+                from mlx_lm.models.vq_switch import _fused
+                eidx = mx.zeros((N,), dtype=mx.uint32)
+                y = _fused(xf, eidx, self.codes[None], self.codebook,
+                           self.vq_scales[None], pack_bits=self.pack_bits)
+            else:
+                from mlx_lm.models.vq_switch import _dense_fused
+                y = _dense_fused(xf, self.codes, self.codebook,
+                                 self.vq_scales, pack_bits=self.pack_bits,
+                                 in_features=self._in_features)
+            return y.astype(x.dtype).reshape(*orig_shape[:-1], OUT)
+        codes = self.codes
+        if self.pack_bits:
+            codes = _unpack_rows(codes, IN // self.codebook.shape[1],
+                                 self.pack_bits)
+        w = _decode(codes, self.codebook.astype(mx.float16),
+                    self.vq_scales, self.group_size, OUT, IN)
+        return (xf @ w.T.astype(x.dtype)).reshape(*orig_shape[:-1], OUT)
+
+
+class VQEmbedding(nn.Module):
+    """Drop-in for nn.Embedding whose table is VQ-coded.
+
+    Decodes ONLY the gathered rows — the whole point of putting VQ on a
+    262k-row table. as_linear() is provided because gemma ties the output
+    head to the input embedding in some configs; it decodes the full table
+    and is deliberately NOT used by the PLE path.
+    """
+
+    def __init__(self, codes, codebook, vq_scales, group_size=64,
+                 pack_bits=0, in_features=None):
+        super().__init__()
+        self.codes = codes
+        self.codebook = codebook
+        self.vq_scales = vq_scales
+        self.group_size = group_size
+        self.pack_bits = pack_bits
+        if pack_bits and in_features is None:
+            raise ValueError("packed codes need explicit in_features")
+        self._in_features = in_features
+        self._k_expect = int(codebook.shape[0])
+        self.freeze()
+
+    @property
+    def num_embeddings(self):
+        return self.codes.shape[0]
+
+    @property
+    def dims(self):
+        return self.codes.shape[1] * self.codebook.shape[1]
+
+    def __call__(self, ids):
+        if int(self.codebook.shape[0]) != self._k_expect:
+            raise RuntimeError("VQ codebook was sharded; must be replicated.")
+        flat = ids.reshape(-1)
+        rows = self.codes[flat]                       # [N, NSUB] or [N, WPR]
+        sc = self.vq_scales[flat]                     # [N, in/G]
+        D = self.codebook.shape[1]
+        if self.pack_bits:
+            NSUB = self._in_features // D
+            rows = _unpack_rows(rows, NSUB, self.pack_bits)
+        N, NSUB = rows.shape
+        IN = NSUB * D
+        w = self.codebook.astype(mx.float16)[rows.reshape(-1)].reshape(N, IN)
+        w = (w.reshape(N, IN // self.group_size, self.group_size)
+             * sc[..., None]).reshape(N, IN)
+        return w.reshape(*ids.shape, IN)
+
+    def as_linear(self, x):
+        OUT = self.codes.shape[0]
+        IN = self.dims
+        w = _decode(self.codes, self.codebook.astype(mx.float16),
+                    self.vq_scales, self.group_size, OUT, IN)
+        return x @ w.T.astype(x.dtype)
+
+
+# ---------------------------------------------------------------------------
+# mlx_lm `model_file` shim. Stock mlx_lm imports Model/ModelArgs from THIS
+# file (it lives inside the checkpoint), so a user needs no local code. We
+# reuse the registry architecture and swap the VQ'd modules for their dense
+# VQ drop-ins BEFORE weights load, so shapes match at load time.
+# ---------------------------------------------------------------------------
+import importlib as _importlib
+import json as _json
+import pathlib as _pathlib
+
+_cfg = _json.load(open(_pathlib.Path(__file__).parent / "config.json"))
+_arch = _importlib.import_module(f"mlx_lm.models.{_cfg['model_type']}")
+ModelArgs = _arch.ModelArgs
+
+
+def _reach(root, path):
+    obj = root
+    parts = path.split(".")
+    for c in parts[:-1]:
+        obj = obj[int(c)] if c.isdigit() else getattr(obj, c)
+    return obj, parts[-1]
+
+
+class Model(_arch.Model):
+    def __init__(self, args):
+        super().__init__(args)
+        for _p, _m in _cfg.get("vq_linear", {}).items():
+            _obj, _leaf = _reach(self, _p)
+            _pb = _m.get("pack_bits", 0)
+            _ct = mx.uint32 if _pb else (mx.uint8 if _m["k"] <= 256 else mx.uint16)
+            _cols = (_m["in"] // _m["dim"] // 32 * _pb) if _pb else _m["in"] // _m["dim"]
+            setattr(_obj, _leaf, VQLinear(
+                mx.zeros((_m["out"], _cols), dtype=_ct),
+                mx.zeros((_m["k"], _m["dim"]), dtype=mx.float16),
+                mx.zeros((_m["out"], _m["in"] // _m["group"]),
+                         dtype=mx.float16),
+                group_size=_m["group"], pack_bits=_pb,
+                in_features=_m["in"] if _pb else None))
+        for _p, _m in _cfg.get("vq_embed", {}).items():
+            _obj, _leaf = _reach(self, _p)
+            _pb = _m.get("pack_bits", 0)
+            _ct = mx.uint32 if _pb else (mx.uint8 if _m["k"] <= 256 else mx.uint16)
+            _cols = (_m["in"] // _m["dim"] // 32 * _pb) if _pb else _m["in"] // _m["dim"]
+            setattr(_obj, _leaf, VQEmbedding(
+                mx.zeros((_m["rows"], _cols), dtype=_ct),
+                mx.zeros((_m["k"], _m["dim"]), dtype=mx.float16),
+                mx.zeros((_m["rows"], _m["in"] // _m["group"]),
+                         dtype=mx.float16),
+                group_size=_m["group"], pack_bits=_pb,
+                in_features=_m["in"] if _pb else None))

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "tool_parser_type": "qwen3_coder",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}