Datasets:
Tasks:
Text Generation
Modalities:
Text
Languages:
English
Size:
< 1K
Tags:
kv-cache
kv-cache-compression
llm-inference
inference-efficiency
efficient-inference
long-context
License:
Add 3B fp16 seed + 7B negative-control cards
Browse files
cards/seed_multimethod_qwen2.5-7b-instruct.json
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"harness_version": "mbe-0.2.0",
|
| 3 |
+
"run_type": "GPU seed run across families (negative control: 4-bit weights)",
|
| 4 |
+
"model": "Qwen/Qwen2.5-7B-Instruct",
|
| 5 |
+
"task": "passkey retrieval (synthetic, single needle)",
|
| 6 |
+
"n_samples": 20,
|
| 7 |
+
"hardware": "cuda / 4-bit weights",
|
| 8 |
+
"note": "Negative control: weights AND KV cache both 4-bit (model loaded in 4-bit to fit a 16GB T4). Double quantization compounds error and collapses retrieval; not a clean KV-quant result. For a clean test use fp16 weights (see the 3B fp16 card).",
|
| 9 |
+
"families": "quantization (KIVI) + eviction (StreamingLLM sink+window, H2O heavy-hitter)",
|
| 10 |
+
"results": [
|
| 11 |
+
{"method": "full", "kv_budget": "100%", "passkey_acc": 1.0},
|
| 12 |
+
{"method": "kivi-4bit", "kv_budget": "25.0%", "passkey_acc": 0.0},
|
| 13 |
+
{"method": "kivi-4bit", "kv_budget": "12.5%", "passkey_acc": 0.0},
|
| 14 |
+
{"method": "streaming", "kv_budget": "25.0%", "passkey_acc": 0.0},
|
| 15 |
+
{"method": "streaming", "kv_budget": "12.5%", "passkey_acc": 0.0},
|
| 16 |
+
{"method": "h2o", "kv_budget": "25.0%", "passkey_acc": 0.0},
|
| 17 |
+
{"method": "h2o", "kv_budget": "12.5%", "passkey_acc": 0.0}
|
| 18 |
+
]
|
| 19 |
+
}
|