Hugging Face's logo

Accio-Lab
/
occamy-1.0-MTP

sglang
mtp
speculative-decoding
draft-head
qwen3_5_moe
Model card Files
xet
 Community
1
occamy-1.0-MTP
File size: 5,019 Bytes
755da9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
{
  "scope": {
    "prompts": 18,
    "repeats": 2,
    "temperature": 0,
    "max_new_tokens": 160,
    "max_concurrent_requests": 1,
    "context_length": 2048,
    "mtp_steps": 1,
    "draft_tokens": 2,
    "cuda_graphs": false,
    "overlap": false,
    "radix_cache": false,
    "deterministic_inference": true,
    "tool_test": "emitted JSON only; no external tool service",
    "code_test": "bounded Python function execution cases",
    "not_covered": [
      "sampled decoding",
      "concurrency",
      "prefix cache",
      "CUDA graphs",
      "MTP2",
      "multimodal MTP",
      "other engines",
      "other quantizations",
      "FP8 with this head",
      "GGUF Q4/Q8 with this head"
    ],
    "authored_prompts": 10,
    "occamy_heldout_prefixes": 8,
    "heldout_prefix_length": 128
  },
  "runtime": {
    "sglang": "0.5.13.post1",
    "source_revision": "85fd90072d1a9f2432842b03588f63b745e524e4",
    "torch": "2.11.0+cu130",
    "hook": "canonical_attention:install"
  },
  "results": {
    "BF16": {
      "baseline": {
        "requests": 36,
        "quality_pass": 12,
        "quality_total": 16,
        "stopped": 22,
        "length_limited": 14,
        "acceptance": null,
        "accepted": 0,
        "proposed": 0,
        "median_e2e_seconds": 4.305030163610354,
        "repeat_token_parity": 18,
        "exact_tokens": 36,
        "exact_logprobs": 36
      },
      "original-head": {
        "requests": 36,
        "quality_pass": 12,
        "quality_total": 16,
        "stopped": 22,
        "length_limited": 14,
        "acceptance": 0.8767942583732058,
        "accepted": 1466,
        "proposed": 1672,
        "median_e2e_seconds": 6.00265346840024,
        "repeat_token_parity": 18,
        "full_token_matches_baseline": 36,
        "median_matched_output_speed_ratio": 0.7472627789451953,
        "exact_tokens": 36,
        "exact_logprobs": 36
      },
      "trained-head": {
        "requests": 36,
        "quality_pass": 12,
        "quality_total": 16,
        "stopped": 22,
        "length_limited": 14,
        "acceptance": 0.8812949640287769,
        "accepted": 1470,
        "proposed": 1668,
        "median_e2e_seconds": 4.158929664408788,
        "repeat_token_parity": 18,
        "full_token_matches_baseline": 36,
        "median_matched_output_speed_ratio": 1.1084637954727476,
        "exact_tokens": 36,
        "exact_logprobs": 36
      }
    },
    "NVFP4": {
      "baseline": {
        "requests": 36,
        "quality_pass": 12,
        "quality_total": 16,
        "stopped": 20,
        "length_limited": 16,
        "acceptance": null,
        "accepted": 0,
        "proposed": 0,
        "median_e2e_seconds": 3.7523877734784037,
        "repeat_token_parity": 18,
        "exact_tokens": 36,
        "exact_logprobs": 36
      },
      "trained-head": {
        "requests": 36,
        "quality_pass": 12,
        "quality_total": 16,
        "stopped": 20,
        "length_limited": 16,
        "acceptance": 0.857630979498861,
        "accepted": 1506,
        "proposed": 1756,
        "median_e2e_seconds": 4.066564766457304,
        "repeat_token_parity": 18,
        "full_token_matches_baseline": 36,
        "median_matched_output_speed_ratio": 1.0389619272561033,
        "exact_tokens": 36,
        "exact_logprobs": 36
      }
    }
  },
  "head_sha256": "56baea0ef88de178c6c858d81adee9e532983798dce81202f21a7483e59001df",
  "acceptance_by_subset": {
    "bf16": {
      "strict": {
        "tokens": 36,
        "logprobs": 36,
        "total": 36
      },
      "variants": {
        "original-head": {
          "authored": {
            "accepted": 338,
            "proposed": 380,
            "rate": 0.8894736842105263
          },
          "occamy_heldout": {
            "accepted": 1128,
            "proposed": 1292,
            "rate": 0.8730650154798761
          }
        },
        "trained-head": {
          "authored": {
            "accepted": 340,
            "proposed": 378,
            "rate": 0.8994708994708994
          },
          "occamy_heldout": {
            "accepted": 1130,
            "proposed": 1290,
            "rate": 0.875968992248062
          }
        }
      }
    },
    "nvfp4": {
      "strict": {
        "tokens": 36,
        "logprobs": 36,
        "total": 36
      },
      "variants": {
        "trained-head": {
          "authored": {
            "accepted": 328,
            "proposed": 384,
            "rate": 0.8541666666666666
          },
          "occamy_heldout": {
            "accepted": 1178,
            "proposed": 1372,
            "rate": 0.858600583090379
          }
        }
      }
    }
  },
  "timing_caveat": "Diagnostic only. Concurrent activity occurred; BF16 original-head timing included additional requests from an invalid parallel attempt. No controlled training-induced speed improvement established. Invalid NVFP4 attempt excluded in full; replacement NVFP4 run used a separate port and checked served model identity."
}