3 people a3c83ff2fd Faster per-token fp8 group quant packed kernel for blackwell (#41326) ...

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>

2026-04-30 18:09:55 -07:00

..

attention

[Bugfix] Fix broken explicit unquantized kv cache dtype support (#38922)

2026-04-09 22:27:53 -07:00

core

[small][batch invariance] Rename the env and internal flags to simplify usage (#26855)

2025-10-16 21:40:25 +00:00

cpu

[Feat] CPU fp8 attn for AMX/AVX-512 (#39445)

2026-04-29 20:43:21 +08:00

cutlass_extensions

[CI/Build] Enable FP8 on NVIDIA Thor (#39712)

2026-04-29 09:48:52 -07:00

libtorch_stable

Faster per-token fp8 group quant packed kernel for blackwell (#41326)

2026-04-30 18:09:55 -07:00

mamba/mamba_ssm

[Bugfix] Use null block (0) for padded block table entries (#35431)

2026-03-30 14:02:51 -07:00

moe

[DeepSeek] Use torch.mm for bf16xbf16->fp32 gemm (#41300)

2026-04-30 16:28:57 -07:00

quantization

[Bugfix] Fix broken explicit unquantized kv cache dtype support (#38922)

2026-04-09 22:27:53 -07:00

quickreduce

[Bugfix][Rocm] fix qr error when different inp shape (#25892)

2025-10-13 10:04:21 -07:00

rocm

[ROCm] Add gfx1102/gfx1103 support (#40037)

2026-04-23 01:32:04 -07:00

activation_kernels.cu

[DSV4] Add silu clamp limit to shared expert (#40950)

2026-04-27 00:37:43 -07:00

async_util.cuh

fused qknorm+rope kernel optimization for SM9.0 (#37376)

2026-04-12 19:58:37 -07:00

cache_kernels_fused.cu

Fuse RoPE and MLA KV-cache write (#25774)

2026-01-09 19:18:37 -08:00

cache_kernels.cu

[Bugfix] Fix IMA in DSA + MTP (#40772)

2026-04-24 01:40:20 -07:00

cache.h

[Perf] Batch KV cache swap copies via cuMemcpyBatchAsync (#38460)

2026-04-03 03:13:23 +00:00

concat_mla_q.cuh

[Attention][Perf][Kernel] Replace torch.cat with vectorized CUDA kernel MLA query concat - DeepSeek-V3.2 (#34917)

2026-03-09 09:50:36 -07:00

cub_helpers.h

[Refactor] Refactor FP8 & INT8 Quant Folder inside w8a8 (#25293)

2025-10-08 10:20:48 -04:00

cuda_compat.h

[Bugfix][ROCm] Fix for warp_size uses on host (#21205)

2025-07-24 00:37:19 -07:00

cuda_utils_kernels.cu

[NVIDIA] Support nvfp4 quantization (#12784)

2025-02-12 19:51:51 -08:00

cuda_utils.h

[Attention] MLA with chunked prefill (#12639)

2025-02-21 15:30:12 -08:00

cuda_vec_utils.cuh

[4/n] Migrate FP4/W4A8 CUTLASS kernels to torch stable ABI (#37503)

2026-03-31 10:21:13 -07:00

cuda_view.cu

[Feature] Support CPU Offloading without Pytorch Pinned Memory that leads to doubled allocation (#32993)

2026-02-13 08:11:26 -08:00

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] fix sleep mode not releasing GPU memory problem on ROCm (#37533)

2026-03-23 06:07:19 -07:00

custom_all_reduce_test.cu

[Distributed] Add custom allreduce support for ROCM (#14125)

2025-03-31 22:49:12 -07:00

custom_all_reduce.cu

[Distributed] Add custom allreduce support for ROCM (#14125)

2025-03-31 22:49:12 -07:00

custom_all_reduce.cuh

[Kernels] Enable Torch Symmetric Memory All-Reduce By Default (#24111)

2025-09-11 09:45:31 -07:00

custom_quickreduce.cu

[Feature] add quick all reduce (#19744)

2025-06-26 20:54:24 -07:00

dispatch_utils.h

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

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

dsv3_fused_a_gemm.cu

[Bugfix] Fix DSV3 kernels breaking _C and _moe_C on unsupported arches (#35123)

2026-02-23 17:11:27 -08:00

fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

[Feat] DeepSeek V4 Rebased (#40860)

2026-04-26 18:31:08 -07:00

fused_qknorm_rope_kernel.cu

fused qknorm+rope kernel optimization for SM9.0 (#37376)

2026-04-12 19:58:37 -07:00

launch_bounds_utils.h

Update launch_bounds_utils.h for correct compile on Multiple Cuda Arch - PTXAS out of range Warning (#25843)

2025-09-30 19:18:19 -07:00

layernorm_kernels.cu

[Feat] DeepSeek V4 Rebased (#40860)

2026-04-26 18:31:08 -07:00

layernorm_quant_kernels.cu

[Feat] DeepSeek V4 Rebased (#40860)

2026-04-26 18:31:08 -07:00

minimax_reduce_rms_kernel.cu

[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

2026-04-11 00:20:20 +08:00

minimax_reduce_rms_kernel.h

[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

2026-04-11 00:20:20 +08:00

nvfp4_kv_cache_kernels.cu

Add nvfp4 support to reshape_and_cache_flash (#37332)

2026-04-17 07:28:00 -07:00

ops.h

[DSV4] Add silu clamp limit to shared expert (#40950)

2026-04-27 00:37:43 -07:00

persistent_topk.cuh

[Feat] DeepSeek V4 Rebased (#40860)

2026-04-26 18:31:08 -07:00

pos_encoding_kernels.cu

[Bugfix] Fix rope (#41113)

2026-04-28 22:42:35 -07:00

sampler.cu

[Feat] DeepSeek V4 Rebased (#40860)

2026-04-26 18:31:08 -07:00

topk.cu

[Bugfix] Fix persistent_topk cooperative deadlock at TopK=1024 (#41189)

2026-04-29 21:03:45 -07:00

torch_bindings.cpp

[DSV4] Add silu clamp limit to shared expert (#40950)

2026-04-27 00:37:43 -07:00

type_convert.cuh

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

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