Xuehao commited on
Commit
a987d08
·
1 Parent(s): 4b15e94

Commit from Azure DevOps update Results

Browse files
Qwen/Qwen3.5-0.8B-autoround-W4A16/results_2026-04-24-14-23-14.json ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/accuracy.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_id": "Qwen/Qwen3.5-0.8B",
3
+ "model_path": "/root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16",
4
+ "scheme": "W4A16",
5
+ "device": "cuda:0",
6
+ "num_gpus": "1",
7
+ "tasks": {
8
+ "piqa": {
9
+ "accuracy": 0.6610446137105549,
10
+ "accuracy_stderr": 0.011044144419710767
11
+ },
12
+ "mmlu": {
13
+ "accuracy": 0.4253667568722404,
14
+ "accuracy_stderr": 0.004073939627807024
15
+ },
16
+ "hellaswag": {
17
+ "accuracy": 0.367257518422625,
18
+ "accuracy_stderr": 0.004810723108378151
19
+ },
20
+ "gsm8k": {
21
+ "accuracy": 0.20697498104624715,
22
+ "accuracy_stderr": 0.011159498164891769
23
+ }
24
+ },
25
+ "status": "success",
26
+ "duration_seconds": "2510.556882251054",
27
+ "eval_framework": "lm_eval+vllm",
28
+ "errors": [],
29
+ "eval_num_gpus": "1"
30
+ }
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/lm_eval_results/__root__.openclaw__workspace__quantized__Qwen_Qwen3.5-0.8B-W4A16/results_2026-04-24T14-15-40.765717.json ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/logs/eval_prompt.txt ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ You are an expert in evaluating quantized LLM models.
2
+ You MUST follow the skill instructions in: /root/.openclaw/workspace/skills/auto_eval/SKILL.md
3
+
4
+ Quantized model path: /root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16
5
+ Runtime artifact directory: /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16
6
+ Raw lm_eval output directory: /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/lm_eval_results
7
+ Evaluation tasks: piqa,mmlu,hellaswag,gsm8k
8
+ Batch size: 8
9
+ Num gpus: 1
10
+
11
+ The quantized model was produced by auto_quant with scheme=W4A16, export_format=auto_round.
12
+ A venv may already exist at /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/venv (created by auto_quant with --system-site-packages).
13
+
14
+ CRITICAL ENVIRONMENT NOTE:
15
+ - System Python has torch+cuda pre-installed. When creating venvs, ALWAYS use:
16
+ python3 -m venv --system-site-packages <path>
17
+ This ensures the venv inherits torch+cuda. Do NOT pip install torch inside the venv.
18
+ - If /root/.venv exists, reuse /root/.venv before creating a new venv.
19
+ - If a venv already exists at /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/venv, reuse it - just install lm_eval and vllm into it.
20
+ - Use uv pip for dependency installation. Prefer:
21
+ uv pip install --python <venv>/bin/python <packages>
22
+ - Do NOT reinstall torch or flash_attn if they already import successfully from the reused environment. Only install them when missing or incompatible.
23
+ - Write evaluation outputs, logs, prompts, copied request/session files, and other runtime artifacts to: /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16
24
+ - When invoking lm_eval, you MUST pass:
25
+ --output_path /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/lm_eval_results
26
+ - Do NOT omit --output_path. Keep the raw lm_eval output files under that exact directory for later upload.
27
+
28
+ IMPORTANT - After evaluation completes, you MUST produce:
29
+
30
+ /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/accuracy.json - evaluation results:
31
+ {
32
+ "model_id": "Qwen/Qwen3.5-0.8B",
33
+ "model_path": "/root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16",
34
+ "scheme": "W4A16",
35
+ "device": "cuda:0",
36
+ "num_gpus": "1",
37
+ "tasks": {
38
+ "<task_name>": {
39
+ "accuracy": <float>,
40
+ "accuracy_stderr": <float or null>
41
+ }
42
+ },
43
+ "status": "success" or "failed",
44
+ "duration_seconds": <float>,
45
+ "eval_framework": "lm_eval+vllm" or "lm_eval+hf" or "manual",
46
+ "errors": [<list of error strings if any>]
47
+ }
48
+
49
+ /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/lm_eval_results/ - raw lm_eval output directory created by:
50
+ lm_eval ... --output_path /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/lm_eval_results
51
+
52
+ The accuracy values MUST be real numbers from actual evaluation runs.
53
+ Write as valid JSON. If evaluation fails, still write accuracy.json with status=failed.
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/logs/quant_prompt.txt ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ You are an expert in LLM quantization using the Intel Auto-Round toolkit.
2
+ You MUST follow the skill instructions in: /root/.openclaw/workspace/skills/auto_quant/SKILL.md
3
+
4
+ Model: Qwen/Qwen3.5-0.8B
5
+ Quantization: W4A16 / RTN
6
+ Export format: auto_round
7
+ Quantized Model Output directory: /root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16
8
+ Runtime artifact directory: /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16
9
+ Runtime device: cuda
10
+ Num gpus: 1
11
+
12
+ Directory responsibilities:
13
+ - Write exported model files to: /root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16
14
+ - Write runtime artifacts such as quant_summary.json, quantize.py, logs, prompts, copied request/session files, and the venv to: /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16
15
+
16
+ CRITICAL SCRIPT REQUIREMENT:
17
+ - Before starting quantization, you MUST first generate the quantization script file:
18
+ /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/quantize.py
19
+ - The file name must be exactly: quantize.py
20
+ - Run quantization by executing that generated quantize.py script
21
+ - Do not use quantize_script.py as the final artifact name
22
+
23
+ CRITICAL ENVIRONMENT NOTE:
24
+ - System Python has torch+cuda pre-installed. When creating venvs, ALWAYS use:
25
+ python3 -m venv --system-site-packages <path>
26
+ This ensures the venv inherits torch+cuda. Do NOT pip install torch inside the venv.
27
+ - If /root/.venv exists, reuse /root/.venv before creating a new venv.
28
+ - Use uv pip for dependency installation. Prefer:
29
+ uv pip install --python <venv>/bin/python <packages>
30
+ - Do NOT reinstall torch or flash_attn if they already import successfully from the reused environment. Only install them when missing or incompatible.
31
+ - This workflow is CUDA-focused. For AutoRound device selection:
32
+ - if Num gpus == 1, prefer device="cuda"
33
+ - if Num gpus > 1, prefer device_map="auto"
34
+ Do NOT default to device_map="0" or device_map="0,1,2,3" unless manual mapping is truly required after auto placement fails.
35
+
36
+ IMPORTANT - After quantization completes (success or failure), you MUST produce:
37
+
38
+ /root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16/quant_summary.json - structured summary:
39
+ {
40
+ "model_id": "Qwen/Qwen3.5-0.8B",
41
+ "scheme": "W4A16",
42
+ "method": "RTN",
43
+ "export_format": "auto_round",
44
+ "device": "cuda",
45
+ "quant_num_gpus": "1",
46
+ "num_gpus": "1",
47
+ "output_dir": "/root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16",
48
+ "runtime_output_dir": "/root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16",
49
+ "quantized_model_dir": "/root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16",
50
+ "status": "success" or "failed",
51
+ "duration_seconds": <float>,
52
+ "original_size_mb": <float or null>,
53
+ "quantized_size_mb": <float or null>,
54
+ "compression_ratio": <float or null>,
55
+ "errors": [<list of error strings>],
56
+ "solutions": [<list of solution strings>],
57
+ "output_files": [<list of file paths in runtime_output_dir>]
58
+ }
59
+
60
+ Write as valid JSON. If quantization fails, still write quant_summary.json with status=failed.
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/quant_summary.json ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/quantize.py ADDED
@@ -0,0 +1,155 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """
3
+ Auto-Round Quantization Script
4
+ Generated by auto_quant skill
5
+
6
+ Model: Qwen/Qwen3.5-0.8B
7
+ Output: /root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16
8
+ Scheme: W4A16
9
+ Iterations: 0 (RTN mode)
10
+ Samples: 128
11
+ Format: auto_round
12
+ Device: cuda (single GPU)
13
+ """
14
+
15
+ import time
16
+ import json
17
+ import os
18
+ import shutil
19
+ from pathlib import Path
20
+
21
+ # Track start time
22
+ start_time = time.time()
23
+
24
+ # Configuration
25
+ model_name_or_path = "Qwen/Qwen3.5-0.8B"
26
+ output_dir = "/root/.openclaw/workspace/quantized/Qwen_Qwen3.5-0.8B-W4A16"
27
+ runtime_output_dir = "/root/.openclaw/workspace/quantized/runs/Qwen_Qwen3.5-0.8B-W4A16"
28
+ scheme = "W4A16"
29
+ iters = 0 # RTN mode
30
+ nsamples = 128
31
+ format_str = "auto_round"
32
+ num_gpus = 1
33
+
34
+ # Track errors and solutions
35
+ errors = []
36
+ solutions = []
37
+
38
+ print(f"Loading model: {model_name_or_path}")
39
+ print(f"Scheme: {scheme}")
40
+ print(f"Iters: {iters} (RTN mode)")
41
+ print(f"nsamples: {nsamples}")
42
+ print(f"Format: {format_str}")
43
+ print(f"Device: cuda")
44
+
45
+ try:
46
+ from auto_round import AutoRound
47
+
48
+ # Create AutoRound instance
49
+ # Single GPU: use device="cuda"
50
+ # Added disable_opt_rtn=True and enable_torch_compile=False to avoid shard writer issues
51
+ ar = AutoRound(
52
+ model_name_or_path,
53
+ scheme=scheme,
54
+ iters=iters,
55
+ nsamples=nsamples,
56
+ device="cuda",
57
+ disable_opt_rtn=True,
58
+ enable_torch_compile=False,
59
+ )
60
+
61
+ # Quantize and save
62
+ print("Starting quantization...")
63
+ ar.quantize_and_save(output_dir=output_dir, format=format_str)
64
+
65
+ print(f"Quantization complete! Output: {output_dir}")
66
+
67
+ # Copy the output to runtime_output_dir as well
68
+ if output_dir != runtime_output_dir:
69
+ print(f"Copying output to runtime directory: {runtime_output_dir}")
70
+ runtime_path = Path(runtime_output_dir)
71
+
72
+ # Copy quantized model files to runtime_output_dir
73
+ quantized_model_src = Path(output_dir)
74
+ for item in quantized_model_src.iterdir():
75
+ dest = runtime_path / item.name
76
+ if item.is_dir():
77
+ if dest.exists():
78
+ shutil.rmtree(dest)
79
+ shutil.copytree(item, dest)
80
+ else:
81
+ shutil.copy2(item, dest)
82
+
83
+ except Exception as e:
84
+ errors.append(str(e))
85
+ print(f"Error during quantization: {e}")
86
+ import traceback
87
+ traceback.print_exc()
88
+
89
+ finally:
90
+ # Calculate duration
91
+ end_time = time.time()
92
+ duration = end_time - start_time
93
+
94
+ # Collect output files
95
+ output_path = Path(runtime_output_dir)
96
+ output_files = []
97
+ original_size_mb = None
98
+ quantized_size_mb = None
99
+
100
+ if output_path.exists():
101
+ for f in sorted(output_path.rglob("*")):
102
+ if f.is_file() and not f.name.endswith(('.pyc', '.pyo', '__pycache__')):
103
+ size = f.stat().st_size
104
+ output_files.append(str(f.relative_to(output_path)))
105
+
106
+ # Try to estimate model sizes
107
+ if 'model.safetensors' in f.name or 'model.bin' in f.name or 'model-00001-of' in f.name:
108
+ # These are typically model weights
109
+ if 'quantized' not in str(f) and original_size_mb is None:
110
+ # This could be original
111
+ pass
112
+
113
+ # Try to get model sizes from output_dir
114
+ model_output_path = Path(output_dir)
115
+ if model_output_path.exists():
116
+ for f in model_output_path.rglob("*"):
117
+ if f.is_file():
118
+ size_mb = f.stat().st_size / (1024 * 1024)
119
+ if 'adapter_model.safetensors' in f.name or 'model-00001-of' in str(f):
120
+ if 'quantized' in str(f) or quantized_size_mb is None:
121
+ quantized_size_mb = size_mb if quantized_size_mb is None else quantized_size_mb + size_mb
122
+ elif 'model.safetensors' in f.name or 'model.bin' in f.name:
123
+ if 'original' not in str(f):
124
+ pass
125
+
126
+ # Build summary
127
+ summary = {
128
+ "model_id": "Qwen/Qwen3.5-0.8B",
129
+ "scheme": scheme,
130
+ "method": "RTN",
131
+ "export_format": format_str,
132
+ "device": "cuda",
133
+ "quant_num_gpus": str(num_gpus),
134
+ "num_gpus": str(num_gpus),
135
+ "output_dir": output_dir,
136
+ "runtime_output_dir": runtime_output_dir,
137
+ "quantized_model_dir": output_dir,
138
+ "status": "success" if len(errors) == 0 else "failed",
139
+ "duration_seconds": duration,
140
+ "original_size_mb": original_size_mb,
141
+ "quantized_size_mb": quantized_size_mb,
142
+ "compression_ratio": (original_size_mb / quantized_size_mb) if (original_size_mb and quantized_size_mb) else None,
143
+ "errors": errors,
144
+ "solutions": solutions,
145
+ "output_files": output_files
146
+ }
147
+
148
+ # Write summary
149
+ summary_path = Path(runtime_output_dir) / "quant_summary.json"
150
+ with open(summary_path, "w") as f:
151
+ json.dump(summary, f, indent=2)
152
+
153
+ print(f"\nSummary written to: {summary_path}")
154
+ print(f"Status: {summary['status']}")
155
+ print(f"Duration: {duration:.2f} seconds")
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/session_eval_313.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/session_eval_313.md ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/session_quant_313.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
Qwen/Qwen3.5-0.8B-autoround-W4A16/run_2026-04-24-14-23-14/session_quant_313.md ADDED
The diff for this file is too large to render. See raw diff