You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Qwen3-Coder MICake C144 Planner (m24)

่ฟ™ไธชไป“ๅบ“ๆ˜ฏ MICake Planner ็š„ๅฝ“ๅ‰ๅ† ๅ†› m24ใ€‚ๅฎƒๆ˜ฏ Qwen3-Coder-30B-A3B-Instruct ไธŠ่ฟž็ปญ LoRA SFT ๅ†ๅˆๅนถ็š„ๅฎŒๆ•ดๆƒ้‡๏ผŒๅฏไปฅ็›ดๆŽฅ็”จ transformers / vLLM / ATOM ๅŠ ่ฝฝใ€‚ๆœฌไป“ๅบ“ไธๆ˜ฏๅ•็‹ฌ็š„ LoRA adapterใ€‚

่ฐฑ็ณป๏ผšQwen3-Coder-30B-A3B-Instruct โ†’ โ€ฆ โ†’ p3๏ผˆๆœฌไป“ๅบ“ๅ…ˆๅ‰ๅ‘ๅธƒ็š„ๆƒ้‡๏ผ‰โ†’ โ€ฆ โ†’ m21 โ†’ m24๏ผˆๆœฌไป“ๅบ“็Žฐๅœจ็š„ๆƒ้‡๏ผ‰ใ€‚m24 ็š„ LoRA ๅŸบๅบงๅ’Œ่ฎญ็ปƒๆ•ฐๆฎ้ƒฝๆฅ่‡ชๅฝ“ๆ—ถ็š„ๅ† ๅ†› m21ใ€‚288 ไธชๆ ทๆœฌใ€640 ่กŒใ€160 ๆญฅใ€2 ไธช epochใ€‚

m24 ๅœจไธŽ่ฎญ็ปƒ้›†ไธ็›ธไบค็š„ held-out ็ป„ๅˆไธŠ๏ผŒ36 ๅฏนไบค้”™ๆต‹่ฏ„็š„ๅนณๅ‡ๅคๆต‹ๅŠ ้€Ÿๆฏ”ๆ˜ฏ 1.1379ร—๏ผŒๅŒๅœบ DSV4.1 Planner ๆ˜ฏ 1.1156ร—ใ€‚ๅทฎ +0.0223๏ผŒ95% ๅŒบ้—ด [+0.0192, +0.0255]๏ผŒp = 0.0001ใ€‚bf16 ๅ’Œ fp8 ๅ„่‡ช 18 ๅฏน๏ผŒไนŸ้ƒฝๆ˜พ่‘—้ซ˜ไบŽ DSV4.1ใ€‚36 ๅฏน้‡Œ m24 ๅ…จ้ƒจ้ซ˜ไบŽๅŒๅœบ DSV4.1ใ€‚

ๆจกๅž‹ๅœจๅšไป€ไนˆ

Planner ไธๅ†™ kernel ๆบ็ ใ€‚ๅฎƒๆไบคไธ€ไปฝ GEMM ้…็ฝฎใ€‚MICake ็”จ่ฟ™ไปฝ้…็ฝฎ็”Ÿๆˆ Triton GEMM๏ผŒๅœจ AMD gfx950 ไธŠ็ผ–่ฏ‘๏ผŒๅšๆญฃ็กฎๆ€งๆฃ€ๆŸฅ๏ผŒๅ†ๆต‹็›ธๅฏนๅŸบ็บฟ้…็ฝฎ็š„ๅ‡ ไฝ•ๅนณๅ‡ๅŠ ้€Ÿๆฏ”ใ€‚ๆฏไธช episode ๆœ€ๅคš 10 ๆฌกๆไบคใ€‚

้…็ฝฎๅˆ†ไธคๆฎต๏ผš

  • schedule๏ผštile_mใ€tile_nใ€tile_kใ€pipeline_depthใ€unroll_factorใ€vectorize_factor
  • realization๏ผšprogram_orderใ€group_mใ€num_warpsใ€num_stagesใ€waves_per_euใ€matrix_instr_nonkdim๏ผˆkpack ๅ›บๅฎšไธบ 1๏ผ‰

็ซฏๅˆฐ็ซฏ่ง’่‰ฒ๏ผšManager ๅ’Œ Critic ๅ›บๅฎšไธบ DSV4.1๏ผŒๅชๆœ‰ Planner ๆขๆจกๅž‹ใ€‚ๅ† ๅ†›ใ€ๆŒ‘ๆˆ˜่€…ใ€DSV4.1 ไธ‰ไธช arm ไบค้”™้‡‡ๆ ทใ€‚held-out ๆฏไธช arm 36 ไธช episode๏ผˆ4 ไธช MICake GPU ร— ๆฏไธช GPU 9 ไธชๆ ทๆœฌ๏ผ‰๏ผŒbf16 ๅ’Œ fp8 ๅ„ 18 ไธชใ€‚

่กจ้‡Œ็š„ๅŠ ้€Ÿๆฏ”ๆ˜ฏๅคๆต‹ๅ€ผ๏ผšagent ๆ‰พๅˆฐ็š„ๆœ€ๅฅฝ้…็ฝฎๅœจๆ–ฐ worker ไธŠๅ†ๆต‹ 2 ๆฌกๅ–ไธญไฝๆ•ฐใ€‚agent ๆฒกๆœ‰ๆไบคๅ‡บๆœ‰ๆ•ˆๆ”น่ฟ›ๆ—ถ่ฎฐ 1.000ร—ใ€‚

่ฆไผ˜ๅŒ–็š„ๅŸบ็บฟ้…็ฝฎ

ๅญ—ๆฎต ๅŸบ็บฟ
tile_m ร— tile_n ร— tile_k 64 ร— 64 ร— 32
pipeline_depth 3
unroll_factor 4
vectorize_factor 4
num_stages 3
num_warps 4
matrix_instr_nonkdim 0
program_order TWO_DIMENSIONAL
group_m 1
waves_per_eu 1
kpack 1

Held-out ๅฝข็Šถ

่ฟ™ 7 ไธชๅฝข็Šถไธๅœจ C144 ่ฎญ็ปƒ้›†้‡Œ๏ผŒไนŸไธๅœจ large ่ฎญ็ปƒ้›†้‡Œใ€‚ไธ€ๆฌกๆไบค็š„ๅˆ†ๆ•ฐๆ˜ฏๅฎƒไปฌ็š„ๅ‡ ไฝ•ๅนณๅ‡ใ€‚่ฎฐๅฝ•้‡Œๆฒกๆœ‰้€ๅฝข็Šถ็š„ speedupใ€‚

workload M N K
gemm.tail-balanced 272 272 288
gemm.tail-rectangular 1025 257 513
gemm.square-4k 4096 4096 4096
gemm.prefill-wide 8192 1024 4096
gemm.prefill-tall 1024 8192 4096
gemm.decode 128 4096 4096
gemm.mlp-7168 4096 7168 2048

Held-out ็ป“ๆžœ

n = 36๏ผŒๅŒไธ€ๆ‰นไปปๅŠกไธŠไบค้”™ๆฏ”่พƒใ€‚m24 ็›ธๅฏนไธŠไธ€ไปฃๅ† ๅ†› m21 ็š„ๅนณๅ‡ best ๅขž็›Šๆ˜ฏ +0.0391๏ผˆ1.1379 ๅฏน 1.0988๏ผ‰๏ผŒๆŒ‰ๆ›ดๅฟซ kernel ๆ™‹ๅ‡ใ€‚

Planner ๅนณๅ‡ๅคๆต‹ ๅˆฐ่พพ 1.01ร— ไธญไฝๆ—ถ้—ด ๅˆฐ่พพ 1.03ร— ไธญไฝๆ—ถ้—ด ๅˆฐ่พพ 1.04ร— ไธญไฝๆ—ถ้—ด ๅˆฐ่พพ 1.05ร— ไธญไฝๆ—ถ้—ด ๆœ€ๅŽไธ€ๆฌกๆ”น่ฟ›็š„ไธญไฝๆ—ถ้—ด ๅข™้’Ÿไธญไฝ
m21 1.0988 93 s 96 s 188 s 284 s 840 s 1019 s
DSV4.1 1.1156 87 s 87 s 87 s 172 s 732 s 919 s
m24 1.1379 93 s 94 s 131 s 267 s 787 s 981 s

ไธ‰ไธช arm ็š„ 36 ไธช episode ้ƒฝๅˆฐ่พพไบ† 1.01ร—ใ€1.03ร—ใ€1.04ร— ๅ’Œ 1.05ร—๏ผŒๅˆฐ่พพ็އ้ƒฝๆ˜ฏ 100%ใ€‚โ€œๆœ€ๅŽไธ€ๆฌกๆ”น่ฟ›โ€ๆ˜ฏ่ฏฅ episode ๆœ€ๅŽไธ€ๆฌกๅˆทๆ–ฐ best ็š„ๆ—ถ้—ด๏ผŒไธ‰ไธช arm ้ƒฝๆ˜ฏ 36/36 ๆœ‰ๅ€ผใ€‚

้€ŸๅบฆไธŠ m24 ็š„ kernel ๆฏ” DSV4.1 ๅฟซ๏ผŒไฝ†ๅˆฐ 1.04ร— ็š„ไธญไฝๆ—ถ้—ดไปๆฏ” DSV4.1 ้•ฟ๏ผˆ131 ็ง’ๅฏน 87 ็ง’๏ผ‰ใ€‚DSV4.1 ็š„็ฌฌไธ€ๆฌกๆœ‰ๆ•ˆๆไบคๅธธๅธธ็›ดๆŽฅ่ฝๅœจ 128ร—128๏ผŒๆ‰€ไปฅ 1.01ร— ๅˆฐ 1.04ร— ๅ‡ ไนŽๅœจๅŒไธ€ๆ—ถๅˆปๅˆฐ่พพใ€‚m24 ็š„้ฆ–ๆไบคๆœ‰ไธ€ๅŠไปๆ˜ฏ 64ร—64๏ผŒ่ฆๅคš่ท‘ๅ‡ ๆฌกๆไบคๆ‰่ƒฝๅˆฐ 128ร—128ใ€‚

dtype ๅฏนๆ•ฐ m24 DSV4.1 ๅทฎ p
bf16 18 1.1327 1.1123 +0.0204 0.0001
fp8 18 1.1431 1.1190 +0.0241 0.0001
ๅˆ่ฎก 36 1.1379 1.1156 +0.0223 0.0001

ๅˆ่ฎก็š„ 95% ๅŒบ้—ดๆ˜ฏ [+0.0192, +0.0255]ใ€‚

้€ๅฏนๅˆ†ๆ•ฐ

dtype ๆ ทๆœฌ m24 DSV4.1 m24 โˆ’ DSV4.1
bf16 0 1.1349 1.1046 +0.0303
bf16 2 1.1243 1.1110 +0.0133
bf16 4 1.1334 1.1199 +0.0135
bf16 6 1.1350 1.1139 +0.0211
bf16 8 1.1262 1.1188 +0.0074
bf16 10 1.1343 1.1038 +0.0305
bf16 12 1.1314 1.1163 +0.0151
bf16 14 1.1328 1.1158 +0.0170
bf16 16 1.1338 1.1165 +0.0173
bf16 18 1.1311 1.1030 +0.0281
bf16 20 1.1333 1.1032 +0.0301
bf16 22 1.1288 1.1094 +0.0194
bf16 24 1.1362 1.1121 +0.0241
bf16 26 1.1400 1.1015 +0.0385
bf16 28 1.1336 1.1143 +0.0193
bf16 30 1.1330 1.1176 +0.0154
bf16 32 1.1307 1.1206 +0.0101
bf16 34 1.1359 1.1186 +0.0173
fp8 1 1.1351 1.1023 +0.0328
fp8 3 1.1395 1.1202 +0.0193
fp8 5 1.1388 1.1214 +0.0174
fp8 7 1.1500 1.1201 +0.0299
fp8 9 1.1431 1.1189 +0.0242
fp8 11 1.1502 1.1183 +0.0319
fp8 13 1.1490 1.1188 +0.0302
fp8 15 1.1372 1.1185 +0.0187
fp8 17 1.1505 1.1213 +0.0292
fp8 19 1.1348 1.1180 +0.0168
fp8 21 1.1366 1.1187 +0.0179
fp8 23 1.1373 1.1211 +0.0162
fp8 25 1.1503 1.1199 +0.0304
fp8 27 1.1398 1.1207 +0.0191
fp8 29 1.1378 1.1191 +0.0187
fp8 31 1.1466 1.1210 +0.0256
fp8 33 1.1480 1.1205 +0.0275
fp8 35 1.1507 1.1224 +0.0283

ๆœ€็ปˆ kernel

36 ๅฏน้‡Œๆœ€็ปˆ้…็ฝฎๆฒกๆœ‰ไธ€ๅฏน็›ธๅŒใ€‚tile ้ƒฝๆ˜ฏ 128ร—128๏ผŒnum_warps ้ƒฝๆ˜ฏ 8๏ผŒๅทฎๅˆซๅœจๅˆซ็š„ๅญ—ๆฎต๏ผš

ๅญ—ๆฎต m24 DSV4.1
unroll_factor ๅ…จ้ƒจๆ˜ฏ 2 ๅ…จ้ƒจๆ˜ฏ 4
pipeline_depth 31/36 ๆ˜ฏ 2๏ผŒ5/36 ๆ˜ฏ 3 35/36 ๆ˜ฏ 3๏ผŒ1/36 ๆ˜ฏ 4
tile_k 24 ๆกๆ˜ฏ 32๏ผŒ12 ๆกๆ˜ฏ 16 32 ๆกๆ˜ฏ 32๏ผŒ4 ๆกๆ˜ฏ 16

ๆŒ‰ dtype ็œ‹ m24 ็š„ tile_k๏ผšbf16 ไธŠ 16/18 ๆกๆ˜ฏ 32ใ€2 ๆกๆ˜ฏ 16๏ผ›fp8 ไธŠ 10/18 ๆกๆ˜ฏ 16ใ€8 ๆกๆ˜ฏ 32ใ€‚vectorize_factor ไธค่พน้ƒฝๆ˜ฏ 4ใ€‚

DSV4.1 ๅœจ 36 ไธช episode ้‡Œไธ€ๆฌก้ƒฝๆฒกๆœ‰ๆไบค่ฟ‡ m24 ็š„้‚ฃไปฝๆœ€็ปˆ้…็ฝฎใ€‚m24 ๆœ‰ 2 ๆฌกๆไบค่ฟ‡ DSV4.1 ็š„ๆœ€็ปˆ้…็ฝฎ๏ผŒๅฝ“ๅœบๆต‹ๅˆฐ 1.1182 ๅ’Œ 1.1191๏ผŒไฝŽไบŽๅฎƒ่‡ชๅทฑ้€‰ๅ‡บ็š„ 1.1404 ๅ’Œ 1.1481ใ€‚

ๅŒไธ€ไปฝ้…็ฝฎ็š„ไธคๆฌกๅคๆต‹็ปๅฏนๅทฎ๏ผŒm24 ไธญไฝๆ•ฐ 0.0015ใ€ๆœ€ๅคง 0.0050๏ผŒDSV4.1 ไธญไฝๆ•ฐ 0.0017ใ€ๆœ€ๅคง 0.0048ใ€‚36 ๅฏน้‡Œ m24 ไธคๆฌกๅคๆต‹็š„่พƒไฝŽๅ€ผ้ƒฝ้ซ˜ไบŽ DSV4.1 ไธคๆฌกๅคๆต‹็š„่พƒ้ซ˜ๅ€ผ๏ผŒๆต‹้‡ๅŒบ้—ดๆฒกๆœ‰้‡ๅ ใ€‚ๅˆ†ๆ•ฐๅทฎๆฅ่‡ชไธŠ้ข่ฟ™ไปฝไธๅŒ็š„้…็ฝฎ๏ผŒไธๆ˜ฏๅŒไธ€ๆฌก kernel ็š„ๆต‹่ฏ•ๆณขๅŠจใ€‚

็ฌฌไธ€ๆฌกๆต‹้€š็š„ๆไบคๆœ‰ 18/36 ๆกๅทฒ็ปๆ˜ฏ 128ร—128ร—32๏ผŒๅฆๅค– 18 ๆกไปไปŽ 64ร—64ร—32 ๅผ€ๅง‹๏ผŒๆœ€ๅŽ้ƒฝ็ˆฌๅˆฐไบ† 128ร—128ใ€‚

m24 ่ฟ™ไธ€่ฝฎ็š„ๆไบค้˜ถๆฎต่ฎกๆ•ฐ๏ผšmeasured 331๏ผŒduplicate 35๏ผŒcompile_fail 26๏ผŒschema_invalid 2๏ผŒverify_fail 1๏ผŒresource_rejected 1ใ€‚

็”จๆณ•

่ฟ™ๆ˜ฏๅˆๅนถๅŽ็š„ Qwen3-MoE ๅ› ๆžœ่ฏญ่จ€ๆจกๅž‹๏ผŒ30.5B ๅ‚ๆ•ฐใ€ๆฏ token ๆฟ€ๆดป็บฆ 3.3Bใ€‚็”จๅŸบๅบงๆจกๅž‹็š„ chat template ๅ’Œๅทฅๅ…ท่ฐƒ็”จๆ ผๅผใ€‚

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Zhangdanyang/qwen3-coder-micake-c144-planner"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, device_map="auto")

ๅŸบๅบง่ฎธๅฏ่ฏๆ˜ฏ Apache 2.0๏ผŒ่งไป“ๅบ“ไธญ็š„ LICENSEใ€‚

Downloads last month
33
Safetensors
Model size
31B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Zhangdanyang/qwen3-coder-micake-c144-planner

Adapter
(58)
this model