3 people 83909ad84b [ROCm][Perf] W4A16: widen skinny GEMM LDS activation loads to b128 (#57015) ...

Signed-off-by: Robert Esclapez Garcia <robert.garcia@amd.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>

2026-10-27 15:46:42 +00:00

..

attention

[ROCm] Octave KV (ROCM_OCTAVE): native 3-bit KV cache backend for AMD GPUs (#59774)

2026-10-26 12:35:00 +08:00

beam

[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and MoE finalize (#62366)

2026-10-21 18:31:57 -07:00

core

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cpu

[CPU] Add GELU-tanh MoE, FP8 checkpoint support and Gemma-4 INT8 TP=4 padding (#62016)

2026-10-23 11:39:37 +08:00

cutlass_extensions

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

gdn

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

kimi_k3

[ROCm][Kimi-K3] Support BF16 KDA recurrent state (#58743)

2026-10-23 18:34:52 +08:00

mamba

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

moe

[Bugfix][MoE] Remap invalid topk_ids in moe_permute without expert_map (#59717)

2026-10-23 16:51:58 -06:00

quantization

[Perf][MXFP4] Optimize activation quant for mxfp4. 3x kernel speedup, 5% throughput improvement (#63101)

2026-10-26 16:51:55 -04:00

rocm

[ROCm][Perf] W4A16: widen skinny GEMM LDS activation loads to b128 (#57015)

2026-10-27 15:46:42 +00:00

activation_kernels.cu

[Refactor] Remove dead code multiple places (#61843)

2026-10-16 15:01:07 +00:00

async_util.cuh

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cache_kernels_fused.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cache_kernels.cu

[GLM 5.3 Perf] Skip scale calculation for concat_and_cache_mla, 1.2% ~ 3.0% E2E TPOT improvement (#62102)

2026-10-24 16:26:19 -04:00

concat_mla_q.cuh

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cooperative_topk_overflow.cuh

[Bugfix] Make cooperative TopK exact on candidate overflow (#59645)

2026-10-20 15:06:55 -07:00

cooperative_topk.cu

[Bugfix] Make cooperative TopK exact on candidate overflow (#59645)

2026-10-20 15:06:55 -07:00

cooperative_topk.cuh

[Perf][Kernel] Size cooperative_topk's single pass by row length, not stride (#62171)

2026-10-23 11:47:29 -04:00

cub_helpers.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cuda_utils_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cuda_utils.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cuda_vec_utils.cuh

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cuda_view.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

cumem_allocator_compat.h

[Core][AMD] Migrate fully transparent sleep mode to ROCm platform (#12695)

2025-11-12 15:24:12 -08:00

cumem_allocator.cpp

[ROCm] Enable the cuMem CUDA-graph pool offload on ROCm (#59523)

2026-10-09 08:29:55 +00:00

custom_all_gather_reduce_scatter_ops.cpp

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

custom_all_gather_reduce_scatter.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

custom_all_gather_reduce_scatter.cuh

[Perf][Distributed] Add low-SM multimem reduce-scatter for SM100/SM103 (#55072)

2026-09-25 03:27:21 +00:00

custom_all_reduce.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

custom_all_reduce.cuh

[Perf][Distributed] Add low-SM multimem reduce-scatter for SM100/SM103 (#55072)

2026-09-25 03:27:21 +00:00

custom_collective_common.cuh

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

dispatch_utils.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

dsv3_fused_a_gemm.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

flashkda_registration.cpp

[Perf] Support internal prefill checkpoints for Mamba prefix caching, 9%~25% TTFT improvement (#52789)

2026-08-22 10:32:40 -07:00

fp32_router_gemm_entry.cu

[Kernel] Add FP32 router GEMM support for Step-5 Preview and Muse Spark Open (#62126)

2026-10-16 22:17:55 -07:00

fp32_router_gemm.cu

[Kernel] Add FP32 router GEMM support for Step-5 Preview and Muse Spark Open (#62126)

2026-10-16 22:17:55 -07:00

fs_io.cpp

[Bugfix][KV Offload] Treat an already-published fs block as stored (#61390)

2026-10-25 17:41:55 +02:00

fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

[Kernel][DSv4.1] Bump FlashMLA to the 2026.09.30 release and its interleaved V4.1 KV cache (#61371)

2026-10-15 13:58:25 -07:00

fused_kimi_k3_mla_key_concat_kv_cache_kernel.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

fused_minimax_m3_qknorm_rope_kv_insert_kernel.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

fused_muse_spark_qknorm_rope_kv_insert_kernel.cu

[Kernel]Add norm fused kernel for Step-5 Preview and Muse Spark Open (#62132)

2026-10-17 01:53:00 -07:00

fused_qknorm_rope_cache_kernel.cu

[Kernel]Add norm fused kernel for Step-5 Preview and Muse Spark Open (#62132)

2026-10-17 01:53:00 -07:00

fused_qknorm_rope_kernel.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

hisparse_kernels.cu

[Perf][HiSparse] Reduce residency hash probes (#61263)

2026-10-26 20:44:15 +00:00

launch_bounds_utils.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

layernorm_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

layernorm_quant_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

minimax_reduce_rms_kernel.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

minimax_reduce_rms_kernel.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

mnnvl_lamport.cuh

[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and MoE finalize (#62366)

2026-10-21 18:31:57 -07:00

muse_spark_mul_tanh_soft_cap_kernel.cu

[Kernel]Add norm fused kernel for Step-5 Preview and Muse Spark Open (#62132)

2026-10-17 01:53:00 -07:00

ngram_embedding_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

nvfp4_ds_mla_cache_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

nvfp4_ds_mla_cache.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

nvfp4_kv_cache_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

ops.h

[Perf][MXFP4] Optimize activation quant for mxfp4. 3x kernel speedup, 5% throughput improvement (#63101)

2026-10-26 16:51:55 -04:00

persistent_topk_overflow.cuh

[2/N][Bugfix] Make persistent TopK exact on candidate overflow (#62936)

2026-10-26 12:25:16 -04:00

persistent_topk.cuh

[Bugfix][Kernel] persistent_topk: sync the CTA around radix barrier arrivals (#63475)

2026-10-26 20:40:32 +00:00

pos_encoding_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

qutlass_registration.cpp

[12/n] final _C library kernel migration (#45415)

2026-06-19 06:57:26 -07:00

sampled_topk.cuh

[Bugfix][Kernel] Fix cooperative_topk illegal memory access on NaN scores (#63118)

2026-10-23 15:28:25 -07:00

sampler.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

spinloop.cpp

[BugFix] Fix clang spinloop mwaitx include (#45532)

2026-07-28 19:00:06 +00:00

topk_histogram_4096_overflow.cuh

[2/N][Bugfix] Make persistent TopK exact on candidate overflow (#62936)

2026-10-26 12:25:16 -04:00

topk_histogram_4096.cuh

[2/N][Bugfix] Make persistent TopK exact on candidate overflow (#62936)

2026-10-26 12:25:16 -04:00

topk.cu

[2/N][Bugfix] Make persistent TopK exact on candidate overflow (#62936)

2026-10-26 12:25:16 -04:00

torch_bindings.cpp

[Perf][MXFP4] Optimize activation quant for mxfp4. 3x kernel speedup, 5% throughput improvement (#63101)

2026-10-26 16:51:55 -04:00

torch_utils.h

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

type_convert.cuh

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00

vocab_parallel_embedding_kernels.cu

[Misc] Consolidate csrc shared headers and prune dead code (#60721)

2026-10-12 09:46:22 -07:00