ray_chan
0 Followers · 0 Following
  • Joined on 2026-07-19
Repositories
1
 Projects Packages Public Activity Starred Repositories
ray_chan created pull request ccf-ai-infra/TileOPs-Metax#44 2026-08-05 23:13:15 +08:00
面向 MetaX C500 的 SwiGLU - fp8 量化融合算子的适配与优化
1b2620103c docs(skills): register-resident GPU memory-hierarchy decision rule
3b980a4f09 refactor(quant_swiglu): drop stage-profiling variants and their scripts
22ba38ceae test(quant_swiglu): enforce bit-exact fp8 contract with torch
d886342ad6 optimize(quant_swiglu): swizzled act_shared layout for nt=128 transposed (-9 to -13% on nt128-transposed)
8eb099ec36 optimize(quant_swiglu): register-resident non-transposed tile, no act_shared (-12.8% sum)
Compare 47 commits »
5df702907a docs(skills): register-resident GPU memory-hierarchy decision rule
da567b4239 refactor(quant_swiglu): drop stage-profiling variants and their scripts
8fe17d8543 test(quant_swiglu): enforce bit-exact fp8 contract with torch
d6feedb525 optimize(quant_swiglu): swizzled act_shared layout for nt=128 transposed (-9 to -13% on nt128-transposed)
1a07fd83bb optimize(quant_swiglu): register-resident non-transposed tile, no act_shared (-12.8% sum)
Compare 12 commits »
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-05 19:57:10 +08:00
5df702907a docs(skills): register-resident GPU memory-hierarchy decision rule
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-05 19:56:39 +08:00
da567b4239 refactor(quant_swiglu): drop stage-profiling variants and their scripts
8fe17d8543 test(quant_swiglu): enforce bit-exact fp8 contract with torch
Compare 2 commits »
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-05 16:39:34 +08:00
d6feedb525 optimize(quant_swiglu): swizzled act_shared layout for nt=128 transposed (-9 to -13% on nt128-transposed)
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-05 14:01:08 +08:00
1a07fd83bb optimize(quant_swiglu): register-resident non-transposed tile, no act_shared (-12.8% sum)
2e196447c8 optimize(quant_swiglu): patch-mapped read + register absmax (5.9% sum)
Compare 2 commits »
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-04 18:45:06 +08:00
832dd9b99c optimize(quant_swiglu): config-tunable tile/thread/stages with full validation
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-04 15:30:27 +08:00
89486149cb optimize(quant_swiglu): hand-written fp32->fp8 pack, bit-exact with torch (1.41x)
ray_chan created branch dev/plot-roofline in ray_chan/TileOPs-Metax 2026-08-04 13:17:54 +08:00
ray_chan pushed to dev/plot-roofline at ray_chan/TileOPs-Metax 2026-08-04 13:17:54 +08:00
074cdf83ef feat: add roofline plot script and C500 profile
b5a5c67072 optimize(quant_swiglu): raise block size to 512 threads (1.42x total)
1b560f4525 optimize(quant_swiglu): fast-math + direct FP8 write (1.33x)
e18a3a1a71 feat(quant_swiglu): add fused SwiGLU + per-channel FP8 cast op
Compare 4 commits »
ray_chan created pull request ccf-ai-infra/TileOPs-Metax#24 2026-08-03 19:56:32 +08:00
[quant_swiglu_fwd_channel_cast_transpose] feat: 新增 spec-only Manifest
5eb069274c fix: update shape rules and add new workload for QuantSwiGLUFwdChannelCastTransposeOp
8498bc6591 add more test cases
a81f77d0b9 manifest: fix roofline flops for quant_swiglu_fwd_channel_cast_transpose
ray_chan pushed to summer-camp-2026 at ray_chan/TileOPs-Metax 2026-08-03 18:18:08 +08:00
3279f324a3 Add draft of manifast
ray_chan pushed to summer-camp-2026 at ray_chan/TileOPs-Metax 2026-08-03 18:13:15 +08:00
3279f324a3 Add draft of manifast