alpharomercoma commited on
Commit
2c6d5df
·
verified ·
1 Parent(s): bfdd4b2

xnnpack 2k: Qwen/Qwen2.5-Math-1.5B-Instruct@aafeb0fc6f22

Browse files
.gitattributes CHANGED
@@ -42,3 +42,4 @@ vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-8k.pte filter=lfs diff=lfs merge=
42
  vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-32k.pte filter=lfs diff=lfs merge=lfs -text
43
  vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-4k.pte filter=lfs diff=lfs merge=lfs -text
44
  xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte filter=lfs diff=lfs merge=lfs -text
 
 
42
  vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-32k.pte filter=lfs diff=lfs merge=lfs -text
43
  vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-4k.pte filter=lfs diff=lfs merge=lfs -text
44
  xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte filter=lfs diff=lfs merge=lfs -text
45
+ xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -25,7 +25,7 @@ Windows in this repo: XNNPACK (CPU) at 2k to 16k; Vulkan (GPU) at 2k to 32k. The
25
 
26
  | Backend | Target | File | Window | Size | Smoke test |
27
  |---|---|---|---|---|---|
28
- | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-2k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-2k.pte) | 2,048 tokens | 1.11 GB | passed ("Paris") |
29
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte) | 4,096 tokens | 1.11 GB | passed ("Paris") |
30
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-8k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-8k.pte) | 8,192 tokens | 1.12 GB | passed ("Paris") |
31
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-16k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-16k.pte) | 16,384 tokens | 1.14 GB | passed ("Paris") |
 
25
 
26
  | Backend | Target | File | Window | Size | Smoke test |
27
  |---|---|---|---|---|---|
28
+ | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte) | 2,048 tokens | 1.11 GB | passed ("Paris") |
29
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte) | 4,096 tokens | 1.11 GB | passed ("Paris") |
30
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-8k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-8k.pte) | 8,192 tokens | 1.12 GB | passed ("Paris") |
31
  | XNNPACK (CPU) | any arm64 | [`xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-16k.pte`](xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-16k.pte) | 16,384 tokens | 1.14 GB | passed ("Paris") |
xnnpack/{Qwen2.5-Math-1.5B-Instruct-8da4w-2k.pte → Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e5613e93c96cca575f0a152e91167c23e99fe658f13e4dd5f7634e34da64e66a
3
  size 1107613952
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:819bbccd30e2cab6c4f59cabbb6bbf7b7f3565cfe3e7742b7c841facc0016262
3
  size 1107613952
xnnpack/config.json CHANGED
@@ -8,12 +8,12 @@
8
  "source_revision": "aafeb0fc6f22cbf0eaeed126eff8be45b0360a35",
9
  "variants": [
10
  {
11
- "file": "Qwen2.5-Math-1.5B-Instruct-8da4w-2k.pte",
12
  "size_bytes": 1107613952,
13
- "sha256": "e5613e93c96cca575f0a152e91167c23e99fe658f13e4dd5f7634e34da64e66a",
14
  "context": 2048,
15
  "source_revision": "aafeb0fc6f22cbf0eaeed126eff8be45b0360a35",
16
- "quantization": "8da4w-g32, int8 embeddings",
17
  "fits_phone_budget": true,
18
  "methods": {
19
  "get_max_context_len": 2048,
 
8
  "source_revision": "aafeb0fc6f22cbf0eaeed126eff8be45b0360a35",
9
  "variants": [
10
  {
11
+ "file": "Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte",
12
  "size_bytes": 1107613952,
13
+ "sha256": "819bbccd30e2cab6c4f59cabbb6bbf7b7f3565cfe3e7742b7c841facc0016262",
14
  "context": 2048,
15
  "source_revision": "aafeb0fc6f22cbf0eaeed126eff8be45b0360a35",
16
+ "quantization": "8da4w-gptq-g32, int8 embeddings, int4 codes solved with GPTQ",
17
  "fits_phone_budget": true,
18
  "methods": {
19
  "get_max_context_len": 2048,
xnnpack/export-report-2k.json CHANGED
@@ -41,12 +41,15 @@
41
  "use_hf_rope": true,
42
  "attention_qkv_bias": true
43
  },
44
- "qmode": "8da4w",
 
45
  "group_size": 32,
46
  "embedding_quantize": "8,0",
 
 
47
  "prefill_chunk": 2048,
48
  "kv_cache_dtype": "fp32",
49
- "label": "8da4w-g32, int8 embeddings",
50
  "description": "ExecuTorch 1.4.0 `export_llm`: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, XNNPACK with extended ops, prefill chunk 2048, fp32 KV cache."
51
  },
52
  "window": {
@@ -101,9 +104,9 @@
101
  },
102
  "files": [
103
  {
104
- "path": "xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-2k.pte",
105
  "bytes": 1107613952,
106
- "sha256": "e5613e93c96cca575f0a152e91167c23e99fe658f13e4dd5f7634e34da64e66a"
107
  }
108
  ],
109
  "estimates": {
@@ -112,16 +115,18 @@
112
  "export_peak_bytes_estimate": 9842653440
113
  },
114
  "host": {
115
- "nproc": 4,
116
- "mem_total_bytes": 16765374464,
117
  "swap_total_bytes": 25769799680,
118
- "cpu_model": "INTEL(R) XEON(R) PLATINUM 8573C",
119
- "peak_rss_convert_bytes": 3942838272,
120
- "peak_rss_export_bytes": 9415229440,
121
- "peak_swap_used_bytes": 4096,
122
- "min_mem_available_bytes": 8141582336,
123
- "export_seconds": 719.8,
124
- "total_seconds": 763.2
 
 
125
  },
126
  "conversion": {
127
  "tensors": 339,
@@ -159,29 +164,142 @@
159
  "smoke": {
160
  "prompt": "<|im_start|>system\nPlease reason step by step, and put your final answer within \\boxed{}.<|im_end|>\n<|im_start|>user\nWhat is the capital of France? Answer with one word.<|im_end|>\n<|im_start|>assistant\n",
161
  "template_error": null,
162
- "reply": "The capital of France is **Paris**. (Note: The capital is actually a city, not a country, but the word \"Paris\" is",
163
- "tokens": 32,
164
  "answered": true,
165
  "answer_required": true,
166
  "passed": true,
167
  "problems": [],
168
  "stats": {
169
- "prefill_token_per_sec": 161.29,
170
- "decode_token_per_sec": 28.2847,
171
  "prompt_tokens": 40,
172
- "generated_tokens": 31,
173
- "model_load_start_ms": 1789303059923,
174
- "model_load_end_ms": 1789303062381,
175
- "inference_start_ms": 1789303062381,
176
- "inference_end_ms": 1789303063725,
177
- "prompt_eval_end_ms": 1789303062629,
178
- "first_token_ms": 1789303062629,
179
- "aggregate_sampling_time_ms": 3,
180
  "SCALING_FACTOR_UNITS_PER_SECOND": 1000
181
  }
182
  },
183
  "run": {
184
- "url": "https://github.com/ExperimentalMachines/executorch-model-exporter/actions/runs/34753863094",
185
- "id": "34753863094"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
186
  }
187
  }
 
41
  "use_hf_rope": true,
42
  "attention_qkv_bias": true
43
  },
44
+ "qmode": "8da4w-gptq",
45
+ "export_llm_qmode": "8da4w",
46
  "group_size": 32,
47
  "embedding_quantize": "8,0",
48
+ "embedding_hqq": false,
49
+ "int4_codes": "gptq",
50
  "prefill_chunk": 2048,
51
  "kv_cache_dtype": "fp32",
52
+ "label": "8da4w-gptq-g32, int8 embeddings, int4 codes solved with GPTQ",
53
  "description": "ExecuTorch 1.4.0 `export_llm`: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, XNNPACK with extended ops, prefill chunk 2048, fp32 KV cache."
54
  },
55
  "window": {
 
104
  },
105
  "files": [
106
  {
107
+ "path": "xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte",
108
  "bytes": 1107613952,
109
+ "sha256": "819bbccd30e2cab6c4f59cabbb6bbf7b7f3565cfe3e7742b7c841facc0016262"
110
  }
111
  ],
112
  "estimates": {
 
115
  "export_peak_bytes_estimate": 9842653440
116
  },
117
  "host": {
118
+ "nproc": 8,
119
+ "mem_total_bytes": 33229729792,
120
  "swap_total_bytes": 25769799680,
121
+ "cpu_model": "AMD EPYC",
122
+ "peak_rss_convert_bytes": 3926851584,
123
+ "peak_rss_export_bytes": 9117868032,
124
+ "peak_swap_used_bytes": 0,
125
+ "min_mem_available_bytes": 22863732736,
126
+ "peak_in_use_bytes": 10365997056,
127
+ "peak_in_use_at": "2026-10-03T23:49:28Z",
128
+ "export_seconds": 312.8,
129
+ "total_seconds": 327.7
130
  },
131
  "conversion": {
132
  "tensors": 339,
 
164
  "smoke": {
165
  "prompt": "<|im_start|>system\nPlease reason step by step, and put your final answer within \\boxed{}.<|im_end|>\n<|im_start|>user\nWhat is the capital of France? Answer with one word.<|im_end|>\n<|im_start|>assistant\n",
166
  "template_error": null,
167
+ "reply": "The capital of France is \"paris\".<|im_end|>",
168
+ "tokens": 10,
169
  "answered": true,
170
  "answer_required": true,
171
  "passed": true,
172
  "problems": [],
173
  "stats": {
174
+ "prefill_token_per_sec": 296.296,
175
+ "decode_token_per_sec": 45.6853,
176
  "prompt_tokens": 40,
177
+ "generated_tokens": 9,
178
+ "model_load_start_ms": 1791071622493,
179
+ "model_load_end_ms": 1791071624343,
180
+ "inference_start_ms": 1791071624344,
181
+ "inference_end_ms": 1791071624676,
182
+ "prompt_eval_end_ms": 1791071624479,
183
+ "first_token_ms": 1791071624479,
184
+ "aggregate_sampling_time_ms": 0,
185
  "SCALING_FACTOR_UNITS_PER_SECOND": 1000
186
  }
187
  },
188
  "run": {
189
+ "url": "https://github.com/ExperimentalMachines/execupack/actions/runs/37158322503",
190
+ "id": "37158322503"
191
+ },
192
+ "measuring_gate": {
193
+ "kind": "measuring_gate",
194
+ "passed": true,
195
+ "prefill_tok_per_sec": 296.3,
196
+ "decode_tok_per_sec": 45.69,
197
+ "prompt_tokens": 40,
198
+ "generated_tokens": 9
199
+ },
200
+ "gate": {
201
+ "passed": true,
202
+ "problems": [],
203
+ "graded": 3,
204
+ "fp32_mean": 0.7674,
205
+ "export_mean": 0.7284,
206
+ "agreed": 3,
207
+ "shortfall": 0.0391,
208
+ "quiet_rows_answered_like_fp32": 4,
209
+ "measures_tool_calling": true,
210
+ "quiet_rows": 4,
211
+ "fp32_undecided_rows": [
212
+ "What is Concord the capital of?",
213
+ "What is Providence the capital of?",
214
+ "What is Augusta the capital of?",
215
+ "What is Nassau the capital of?",
216
+ "Who is the author of Mercy?",
217
+ "Who is the author of Landfall?",
218
+ "Who is the author of Quicksand?",
219
+ "Who is the author of Nightfall?",
220
+ "In what country is Linkoping?"
221
+ ],
222
+ "rows": {
223
+ "What is Concord the capital of?": {
224
+ "fp32": 0.3837,
225
+ "export": 0.3814,
226
+ "kind": "search"
227
+ },
228
+ "What is Providence the capital of?": {
229
+ "fp32": 0.0733,
230
+ "export": 0.1264,
231
+ "kind": "search"
232
+ },
233
+ "What is Augusta the capital of?": {
234
+ "fp32": 0.4385,
235
+ "export": 0.3479,
236
+ "kind": "search"
237
+ },
238
+ "What is Nassau the capital of?": {
239
+ "fp32": 0.2163,
240
+ "export": 0.0551,
241
+ "kind": "search"
242
+ },
243
+ "Who is the author of Mercy?": {
244
+ "fp32": 0.3152,
245
+ "export": 0.3211,
246
+ "kind": "search"
247
+ },
248
+ "Who is the author of Landfall?": {
249
+ "fp32": 0.2097,
250
+ "export": 0.2282,
251
+ "kind": "search"
252
+ },
253
+ "Who is the author of Quicksand?": {
254
+ "fp32": 0.2175,
255
+ "export": 0.1711,
256
+ "kind": "search"
257
+ },
258
+ "Who is the author of Nightfall?": {
259
+ "fp32": 0.2885,
260
+ "export": 0.2729,
261
+ "kind": "search"
262
+ },
263
+ "In what country is Ostrava?": {
264
+ "fp32": 0.8516,
265
+ "export": 0.8452,
266
+ "kind": "search"
267
+ },
268
+ "In what country is Gdynia?": {
269
+ "fp32": 0.6144,
270
+ "export": 0.5423,
271
+ "kind": "search"
272
+ },
273
+ "In what country is Rijeka?": {
274
+ "fp32": 0.8363,
275
+ "export": 0.7976,
276
+ "kind": "search"
277
+ },
278
+ "In what country is Linkoping?": {
279
+ "fp32": 0.0505,
280
+ "export": 0.0311,
281
+ "kind": "search"
282
+ },
283
+ "What is the capital of Norway?": {
284
+ "fp32": 0.2098,
285
+ "export": 0.3573,
286
+ "kind": "quiet"
287
+ },
288
+ "What is 8 times 7?": {
289
+ "fp32": 0.2836,
290
+ "export": 0.4076,
291
+ "kind": "quiet"
292
+ },
293
+ "Who wrote Pride and Prejudice?": {
294
+ "fp32": 0.0631,
295
+ "export": 0.0868,
296
+ "kind": "quiet"
297
+ },
298
+ "What is the chemical symbol for iron?": {
299
+ "fp32": 0.4132,
300
+ "export": 0.5774,
301
+ "kind": "quiet"
302
+ }
303
+ }
304
  }
305
  }