Upload folder using huggingface_hub
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- rows/bonsai-0p5b__awq__wikitext2__n128__s0__T4.json +37 -0
- rows/bonsai-0p5b__awq__wikitext2__n128__s1__T4.json +37 -0
- rows/bonsai-0p5b__awq__wikitext2__n128__s2__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s0__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s0__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s1__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s1__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s2__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n128__s2__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s0__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s0__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s1__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s1__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s2__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n32__s2__T4.json +37 -0
- rows/qwen25-3b__awq__openhermes__n512__s0__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n512__s1__A10.json +37 -0
- rows/qwen25-3b__awq__openhermes__n512__s2__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s0__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s0__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s1__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s1__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s2__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n128__s2__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s0__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s0__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s1__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s1__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s2__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n32__s2__T4.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n512__s0__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n512__s1__A10.json +37 -0
- rows/qwen25-3b__awq__wikitext2__n512__s2__A10.json +37 -0
- rows/qwen25-3b__fp16-awq__none__n0__s0__A10.json +37 -0
- rows/qwen25-3b__fp16-awq__none__n0__s0__T4.json +37 -0
- rows/qwen25-3b__fp16-gptq__none__n0__s0__A10.json +37 -0
- rows/qwen25-3b__fp16-gptq__none__n0__s0__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s0__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s0__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s1__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s1__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s2__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n128__s2__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s0__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s0__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s1__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s1__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s2__A10.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n32__s2__T4.json +37 -0
- rows/qwen25-3b__gptq__openhermes__n512__s0__A10.json +37 -0
rows/bonsai-0p5b__awq__wikitext2__n128__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "deepgrove/Bonsai",
|
| 4 |
+
"params": "0.5B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "",
|
| 33 |
+
"transformers_version": "",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OSError(\"/cache/quants/bonsai-0p5b__awq__wikitext2__n128__s0 does not appear to have a file named modeling_qllama.py. Checkout 'https://huggingface.co//cache/quants/bonsai-0p5b__awq__wikitext2__n128__s0/tree/main' for available files.\")",
|
| 36 |
+
"_row_id": "bonsai-0p5b__awq__wikitext2__n128__s0__T4"
|
| 37 |
+
}
|
rows/bonsai-0p5b__awq__wikitext2__n128__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "deepgrove/Bonsai",
|
| 4 |
+
"params": "0.5B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "",
|
| 33 |
+
"transformers_version": "",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OSError(\"/cache/quants/bonsai-0p5b__awq__wikitext2__n128__s1 does not appear to have a file named modeling_qllama.py. Checkout 'https://huggingface.co//cache/quants/bonsai-0p5b__awq__wikitext2__n128__s1/tree/main' for available files.\")",
|
| 36 |
+
"_row_id": "bonsai-0p5b__awq__wikitext2__n128__s1__T4"
|
| 37 |
+
}
|
rows/bonsai-0p5b__awq__wikitext2__n128__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "deepgrove/Bonsai",
|
| 4 |
+
"params": "0.5B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "",
|
| 33 |
+
"transformers_version": "",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OSError(\"/cache/quants/bonsai-0p5b__awq__wikitext2__n128__s2 does not appear to have a file named modeling_qllama.py. Checkout 'https://huggingface.co//cache/quants/bonsai-0p5b__awq__wikitext2__n128__s2/tree/main' for available files.\")",
|
| 36 |
+
"_row_id": "bonsai-0p5b__awq__wikitext2__n128__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6539,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.18,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.2064,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1255,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 631.5,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6145,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.44,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.7397,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9722,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6121,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.14,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1646,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0791,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 637.8,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5762,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.44,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.7014,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9223,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6408,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.15,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1933,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1345,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 633.6,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n128__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5987,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.44,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.7239,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9824,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n128__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6582,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.16,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.2107,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1373,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 459.7,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6207,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.47,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.7459,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9767,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6013,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.14,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1538,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1076,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 433.3,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5626,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.45,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.6878,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9602,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6236,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.15,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1761,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1484,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 455.6,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n32__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5782,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.43,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__openhermes__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.7034,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9943,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n32__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n512__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=512 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.96 GiB. GPU 0 has a total capacity of 22.06 GiB of which 419.44 MiB is free. Process 1 has 21.64 GiB memory in use. Of the allocated memory 16.42 GiB is allocated by PyTorch, and 4.92 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid ",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n512__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n512__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=512 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.87 GiB. GPU 0 has a total capacity of 22.06 GiB of which 4.56 GiB is free. Process 1 has 17.49 GiB memory in use. Of the allocated memory 16.76 GiB is allocated by PyTorch, and 436.95 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid ",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n512__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__openhermes__n512__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=openhermes n=512 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.91 GiB. GPU 0 has a total capacity of 22.06 GiB of which 3.35 GiB is free. Process 1 has 18.70 GiB memory in use. Of the allocated memory 16.86 GiB is allocated by PyTorch, and 1.54 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fr",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__openhermes__n512__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5527,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.17,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1052,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1833,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 635.4,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5067,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.6319,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 4.0462,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5184,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.22,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.0709,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0834,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 615.6,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4903,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.6155,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9314,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5045,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.14,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.057,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1382,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 625.0,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n128__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4714,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.5966,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9895,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n128__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4937,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.17,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.0462,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1792,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 462.2,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4563,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.5815,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 4.0172,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5522,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.21,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.1047,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0982,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 458.3,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5297,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.6549,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.9353,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4826,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 14.19,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": 9.0351,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.1126,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 456.5,
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n32__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.4501,
|
| 10 |
+
"vram_gb": 5.42,
|
| 11 |
+
"tokens_per_sec": 4.65,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__awq__wikitext2__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": 8.5753,
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": 3.948,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "gemm",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n32__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n512__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=512 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.72 GiB. GPU 0 has a total capacity of 22.06 GiB of which 1.27 GiB is free. Process 1 has 20.78 GiB memory in use. Of the allocated memory 15.79 GiB is allocated by PyTorch, and 4.69 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fr",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n512__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n512__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=512 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.71 GiB. GPU 0 has a total capacity of 22.06 GiB of which 427.44 MiB is free. Process 1 has 21.63 GiB memory in use. Of the allocated memory 16.48 GiB is allocated by PyTorch, and 4.85 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid ",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n512__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__awq__wikitext2__n512__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "awq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "wikitext2-train",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": "",
|
| 10 |
+
"vram_gb": "",
|
| 11 |
+
"tokens_per_sec": "",
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "AWQ zero_point g128 GEMM calib_len=512; autoawq 0.2.9 (upstream-deprecated; pinned to its last-tested torch2.6.0/transformers4.51.3 combo); calib=wikitext2 n=512 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "failed",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 512,
|
| 21 |
+
"ppl_fp16_wt2": "",
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": "",
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "OutOfMemoryError('CUDA out of memory. Tried to allocate 4.62 GiB. GPU 0 has a total capacity of 22.06 GiB of which 411.44 MiB is free. Process 1 has 21.65 GiB memory in use. Of the allocated memory 17.10 GiB is allocated by PyTorch, and 4.25 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid ",
|
| 36 |
+
"_row_id": "qwen25-3b__awq__wikitext2__n512__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__fp16-awq__none__n0__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "none-fp16",
|
| 6 |
+
"bits": 16,
|
| 7 |
+
"calibration_set": "none",
|
| 8 |
+
"calibration_size": 0,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.0,
|
| 10 |
+
"vram_gb": 9.49,
|
| 11 |
+
"tokens_per_sec": 27.82,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "fp16 reference row for the awq stack; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": 8.5525,
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "hf-fp16",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__fp16-awq__none__n0__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__fp16-awq__none__n0__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "none-fp16",
|
| 6 |
+
"bits": 16,
|
| 7 |
+
"calibration_set": "none",
|
| 8 |
+
"calibration_size": 0,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.0,
|
| 10 |
+
"vram_gb": 9.49,
|
| 11 |
+
"tokens_per_sec": 24.86,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "fp16 reference row for the awq stack; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": 8.1252,
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": 3.7438,
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "hf-fp16",
|
| 32 |
+
"torch_version": "2.6.0+cu124",
|
| 33 |
+
"transformers_version": "4.51.3",
|
| 34 |
+
"quantlib": "autoawq-0.2.9",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__fp16-awq__none__n0__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__fp16-gptq__none__n0__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "none-fp16",
|
| 6 |
+
"bits": 16,
|
| 7 |
+
"calibration_set": "none",
|
| 8 |
+
"calibration_size": 0,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.0,
|
| 10 |
+
"vram_gb": 9.49,
|
| 11 |
+
"tokens_per_sec": 34.78,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "fp16 reference row for the gptq stack; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "hf-fp16",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__fp16-gptq__none__n0__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__fp16-gptq__none__n0__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "none-fp16",
|
| 6 |
+
"bits": 16,
|
| 7 |
+
"calibration_set": "none",
|
| 8 |
+
"calibration_size": 0,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.0,
|
| 10 |
+
"vram_gb": 9.49,
|
| 11 |
+
"tokens_per_sec": 18.55,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "fp16 reference row for the gptq stack; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": "",
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": "",
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": "",
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": "",
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "hf-fp16",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__fp16-gptq__none__n0__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.657,
|
| 10 |
+
"vram_gb": 14.37,
|
| 11 |
+
"tokens_per_sec": 16.56,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.2094,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0304,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 582.2,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5998,
|
| 10 |
+
"vram_gb": 11.34,
|
| 11 |
+
"tokens_per_sec": 2.97,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.7252,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.8578,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6619,
|
| 10 |
+
"vram_gb": 14.37,
|
| 11 |
+
"tokens_per_sec": 16.62,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.2143,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0429,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 577.1,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6128,
|
| 10 |
+
"vram_gb": 7.49,
|
| 11 |
+
"tokens_per_sec": 4.28,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.7382,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.8838,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6432,
|
| 10 |
+
"vram_gb": 12.45,
|
| 11 |
+
"tokens_per_sec": 16.0,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.1956,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0604,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 598.2,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n128__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 128,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.5778,
|
| 10 |
+
"vram_gb": 9.41,
|
| 11 |
+
"tokens_per_sec": 4.04,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=128 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n128__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.7032,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.8897,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n128__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.8183,
|
| 10 |
+
"vram_gb": 11.69,
|
| 11 |
+
"tokens_per_sec": 16.25,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.3707,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0696,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 402.1,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s0__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s0__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.7669,
|
| 10 |
+
"vram_gb": 7.49,
|
| 11 |
+
"tokens_per_sec": 4.0,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.8923,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.8965,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s0__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s1__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.7619,
|
| 10 |
+
"vram_gb": 13.62,
|
| 11 |
+
"tokens_per_sec": 16.1,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.3143,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.099,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 390.6,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s1__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s1__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.7086,
|
| 10 |
+
"vram_gb": 9.41,
|
| 11 |
+
"tokens_per_sec": 3.79,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=1; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s1 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 1,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.834,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.9325,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s1__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s2__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6978,
|
| 10 |
+
"vram_gb": 13.62,
|
| 11 |
+
"tokens_per_sec": 16.3,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.2502,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0731,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 397.6,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s2__A10"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n32__s2__T4.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 32,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6379,
|
| 10 |
+
"vram_gb": 11.34,
|
| 11 |
+
"tokens_per_sec": 3.68,
|
| 12 |
+
"gpu_tier": "T4",
|
| 13 |
+
"colab_fit": "yes",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-08-01",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=32 seed=2; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n32__s2 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 2,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.1254,
|
| 22 |
+
"ppl_quant_wt2": 8.7633,
|
| 23 |
+
"ppl_fp16_chat": 3.7439,
|
| 24 |
+
"ppl_quant_chat": 3.8894,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": "",
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n32__s2__T4"
|
| 37 |
+
}
|
rows/qwen25-3b__gptq__openhermes__n512__s0__A10.json
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "text-generation",
|
| 3 |
+
"model": "Qwen/Qwen2.5-3B-Instruct",
|
| 4 |
+
"params": "3B",
|
| 5 |
+
"quant_method": "gptq",
|
| 6 |
+
"bits": 4,
|
| 7 |
+
"calibration_set": "openhermes-2.5",
|
| 8 |
+
"calibration_size": 512,
|
| 9 |
+
"ppl_delta_vs_fp16": 0.6588,
|
| 10 |
+
"vram_gb": 17.38,
|
| 11 |
+
"tokens_per_sec": 15.91,
|
| 12 |
+
"gpu_tier": "A10",
|
| 13 |
+
"colab_fit": "n/a",
|
| 14 |
+
"webgpu_runnable": "[UNVERIFIED]",
|
| 15 |
+
"hf_link": "",
|
| 16 |
+
"verified_date": "2026-07-31",
|
| 17 |
+
"notes": "GPTQ g128 desc_act=True damp=0.01 calib_len=2048; calib=openhermes n=512 seed=0; ppl: wikitext2-test + OpenHermes-heldout, ctx2048 non-overlapping, token-capped per tier; tok/s: greedy b1 64->256 median-of-3 (incl. prefill); Qwen Research licence (card license: other), not Apache-2.0 like the 1.5B \u2014 quantized weights NOT auto-published, rows only, pending licence review; artifact modal vol quant-bench:/quants/qwen25-3b__gptq__openhermes__n512__s0 (not pushed to HF yet)",
|
| 18 |
+
"status": "ok",
|
| 19 |
+
"seed": 0,
|
| 20 |
+
"calib_seq_len": 2048,
|
| 21 |
+
"ppl_fp16_wt2": 8.5524,
|
| 22 |
+
"ppl_quant_wt2": 9.2112,
|
| 23 |
+
"ppl_fp16_chat": 3.8919,
|
| 24 |
+
"ppl_quant_chat": 4.0624,
|
| 25 |
+
"ppl_ctx_len": 2048,
|
| 26 |
+
"batch_size": 1,
|
| 27 |
+
"prompt_tokens": 64,
|
| 28 |
+
"gen_tokens": 256,
|
| 29 |
+
"artifact_gb": 2.08,
|
| 30 |
+
"quant_time_s": 1453.5,
|
| 31 |
+
"load_backend": "torch-fallback",
|
| 32 |
+
"torch_version": "2.8.0+cu128",
|
| 33 |
+
"transformers_version": "5.14.1",
|
| 34 |
+
"quantlib": "gptqmodel-7.3.2",
|
| 35 |
+
"error": "",
|
| 36 |
+
"_row_id": "qwen25-3b__gptq__openhermes__n512__s0__A10"
|
| 37 |
+
}
|