c0de128 ac23d0ba18 [Bugfix][Hardware][AMD] Use dynamic WARP_SIZE in sampler vectorized_process ...

Replace hardcoded WARP_SIZE=32 with the dynamic WARP_SIZE macro from
cuda_compat.h to correctly support both Wave64 (MI300X/gfx942) and
Wave32 (Strix Halo/gfx1151) architectures.

The previous hardcoded value was incorrect for AMD CDNA GPUs which use
64-wide wavefronts. While the current static_assert (kWarpSize >= 4)
passes for both 32 and 64, having inconsistent WARP_SIZE definitions
across the codebase is a maintenance issue and potential latent bug.

Changes:
- Add cuda_compat.h include for WARP_SIZE macro
- Replace local WARP_SIZE constant with kWarpSize from cuda_compat.h
- Update static_assert and comments to use kWarpSize

Signed-off-by: c0de128 <kevin.mckay@outlook.com>

2025-12-24 09:02:06 -06:00

..

attention

[Performance][MLA][ROCm] Remove redundant D2D copy in deepseek (#27457)

2025-11-26 12:45:28 +08:00

core

…

cpu

[CPU][Bugfix] Fix ppc64le CPU build (#30871)

2025-12-19 12:26:35 +00:00

cutlass_extensions

…

mamba/mamba_ssm

[V1] [Hybrid] Mamba1 Automatic Prefix Caching (#26377)

2025-11-02 04:16:23 -08:00

moe

[Kernel][Quantization][MoE] add marlin kernel support for turing (sm75) (#29901)

2025-12-16 14:35:28 -08:00

quantization

[FIX] FP4 quantization kernel padding initialization bug (#31097)

2025-12-23 08:45:18 -08:00

quickreduce

…

rocm

Improve wvsplitK tile and balance heristics. (#29937)

2025-12-09 23:51:32 +00:00

sparse/cutlass

…

activation_kernels.cu

[Perf][Kernels] Vectorize csrc/activations_kernels.cu (#29512)

2025-12-16 14:56:02 -08:00

cache_kernels.cu

[DeepSeek v3.2] Remove unnecessary syncwarps (#31047)

2025-12-23 21:33:30 -08:00

cache.h

[Misc] Remove unused custom ops copy_blocks and copy_blocks_mla (#30967)

2025-12-23 18:22:35 -08:00

cub_helpers.h

…

cuda_compat.h

…

cuda_utils_kernels.cu

…

cuda_utils.h

…

cuda_view.cu

Simplify from_blob usage in get_cuda_view_from_cpu_tensor (#29027)

2025-11-22 10:35:32 +00:00

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

[NIXL][Bugfix] Fix NIXL/RDMA registration failure over CuMemAllocator (#29569)

2025-12-17 01:52:58 -08:00

custom_all_reduce_test.cu

…

custom_all_reduce.cu

…

custom_all_reduce.cuh

…

custom_quickreduce.cu

…

dispatch_utils.h

[Performance] Fused blockwise quant RMS norm (#27883)

2025-12-07 16:38:04 +00:00

fused_qknorm_rope_kernel.cu

[ROCm] [Critical]: Remove unused variable (#31156)

2025-12-22 08:28:22 -08:00

launch_bounds_utils.h

…

layernorm_kernels.cu

[kernel][perf] support uncontiguous input for rms_norm kernel (#28103)

2025-11-20 19:39:09 -08:00

layernorm_quant_kernels.cu

[Kernel] Optimize rms_norm kernel (#27931)

2025-11-11 18:02:23 +00:00

ops.h

[Kernel]Support W4A8 Grouped GEMM on Hopper (#29691)

2025-12-08 19:29:06 -08:00

permute_cols.cu

…

pos_encoding_kernels.cu

…

sampler.cu

[Bugfix][Hardware][AMD] Use dynamic WARP_SIZE in sampler vectorized_process

2025-12-24 09:02:06 -06:00

torch_bindings.cpp

[Misc] Remove unused custom ops copy_blocks and copy_blocks_mla (#30967)

2025-12-23 18:22:35 -08:00

type_convert.cuh

[ROCm] [Bugfix] Fix fused_qknorm_rope_kernel rocm compatibility (#28500)

2025-11-12 05:01:14 -08:00