Issues

[Config] bf16_tuned_gemm.csv ships no gfx942 rows - every bf16 GEMM falls back to torch on MI300X

#5380

· skimanwhite opened

on Sep 9, 2026

[tune] a4w4_blockscale tuner has no --libtype axis, so CK Tile's MX GEMM can never compete

#5356

· ThomasNing opened

on Sep 8, 2026

`get_num_xcds()` is hard-coded to 8 everywhere

#5344

· neoblizz opened

on Sep 8, 2026

[MI355X][MegaMoEV2] GLM-5.2 EP8 tuning handoff and C14 performance gap

#5322

· jiejingzhangamd opened

on Sep 8, 2026

[Bug] AITER_FLYDSL_FORCE_REDUCE: comment states the inverse default, and the switch is inert on the tuned-config path

#5299

· stefanskiasan opened

on Sep 7, 2026

deepgemm_fp8_paged_mqa_logits crashes with SIGFPE for ChunkK=64 on gfx950

#5298

· stefanskiasan opened

on Sep 6, 2026

[Perf] pa_decode_sparse: BLOCK_K=64 is suboptimal above the CU count — up to +43 % on gfx950 with a load-dependent choice

#5293

· stefanskiasan opened

on Sep 5, 2026

mhc_post silently computes wrong results when post_layer_mix / comb_res_mix are non-contiguous

#5292

· Duyi-Wang opened

on Sep 5, 2026

[Issue]: gemm_a8w8_blockscale (Triton) returns NaN when K/NUM_KSPLIT is not a multiple of BLOCK_SIZE_K (gfx1201 DEFAULT config, K=12800 -> vLLM garbage on R9700)

#5286

· xy7ra opened

on Sep 5, 2026

[Perf] a8w8_blockscale tuned table for gfx950 appears to be generated without --splitK: re-tuning gains 13-61% on low-M shapes

#5277

· stefanskiasan opened

on Sep 4, 2026

Untuned-shape logging exists only for bf16 (AITER_TUNE_GEMM); fp8/blockscale/a4w4 have none

#5267

· ThomasNing opened

on Sep 3, 2026

JIT baton dead-holder check is ambiguous under --network host (pid + hostname collide across containers)

#5266

· ThomasNing opened

on Sep 3, 2026

Navigation Menu

Uh oh!

All issues

[Config] bf16_tuned_gemm.csv ships no gfx942 rows - every bf16 GEMM falls back to torch on MI300X

[tune] a4w4_blockscale tuner has no --libtype axis, so CK Tile's MX GEMM can never compete

`get_num_xcds()` is hard-coded to 8 everywhere

[MI355X][MegaMoEV2] GLM-5.2 EP8 tuning handoff and C14 performance gap

[Bug] AITER_FLYDSL_FORCE_REDUCE: comment states the inverse default, and the switch is inert on the tuned-config path

deepgemm_fp8_paged_mqa_logits crashes with SIGFPE for ChunkK=64 on gfx950

[Perf] pa_decode_sparse: BLOCK_K=64 is suboptimal above the CU count — up to +43 % on gfx950 with a load-dependent choice

mhc_post silently computes wrong results when post_layer_mix / comb_res_mix are non-contiguous

[Issue]: gemm_a8w8_blockscale (Triton) returns NaN when K/NUM_KSPLIT is not a multiple of BLOCK_SIZE_K (gfx1201 DEFAULT config, K=12800 -> vLLM garbage on R9700)

[Perf] a8w8_blockscale tuned table for gfx950 appears to be generated without --splitK: re-tuning gains 13-61% on low-M shapes

Untuned-shape logging exists only for bf16 (AITER_TUNE_GEMM); fp8/blockscale/a4w4 have none

JIT baton dead-holder check is ambiguous under --network host (pid + hostname collide across containers)

Uh oh!

Search results