File size: 1,282 Bytes
0b73719
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
{
  "quantization_method": "autoround",
  "quantization_tool": "auto_round",
  "output_format": "auto_round",
  "strategy": "mixed_precision_4bit_8bit_with_calibration",
  "based_on": "quantize_autoround_simple_fast_non_prod_only.py + quantize_autoround_4bit_hf_dataset.py",
  "dataset_source": "same_as_layer_observer",
  "mixed_precision": {
    "expert_bits": 4,
    "default_bits": 8,
    "description": "4-bit for MoE experts, 8-bit for attention/FFN, FP16 for router/norms/embeddings"
  },
  "group_size": 128,
  "symmetric": true,
  "vllm_compatible": "check_docs_for_mixed_precision_support",
  "source_model": "artifacts/GLM-4.7/agent_calibration_mix_v5.jsonl/pruned_models/reap-seed_42-0.25",
  "calibration": {
    "samples": 8192,
    "seq_len": 2048,
    "dataset_name": "artifacts/agent_calibration_mix_v6.jsonl",
    "dataset_config_file": null,
    "seed": 42
  },
  "optimization_iters": 200,
  "layer_stats": {
    "ignored_fp16": 268,
    "expert_4bit": 31773,
    "other_8bit": 377,
    "total_quantized": 32150
  },
  "ignore_patterns": [
    ".*embed_tokens.*",
    ".*lm_head.*",
    ".*shared_experts.*",
    ".*shared_head.*",
    ".*router.*",
    ".*ffn_gate_inp.*",
    ".*norm.*\\.weight$",
    ".*layernorm.*",
    ".*rmsnorm.*",
    ".*\\.bias$"
  ]
}