Jee Jee LiandGitHub 56afa45bf7 Fuse qknorm+rope+insert (#20) ...

* Init

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* Delete comments

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* NHD layout

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* torch stable abi

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* torch stable abi

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* FMT

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

* FMT

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

---------

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

2026-06-03 23:10:47 +08:00

..

attention

[9/n] Migrate attention and cache kernels to torch stable ABI (continued) (#43717)

2026-05-29 04:44:45 +00:00

core

[5/n] Migrate CUTLASS MLA, hadamard, awq, allspark and DSV3 fused a gemm to torch stable ABI (continued) (#42339)

2026-05-13 07:24:39 +00:00

cpu

[CPU] Enable non-divisible GQA for decode workitems in mixed batches (#43032)

2026-05-26 14:15:47 +08:00

cutlass_extensions

[6/n] Migrate activation kernels, gptq, gguf, non cutlass w8a8 to libtorch stable ABI (continued) (#42663)

2026-05-20 00:18:12 -07:00

libtorch_stable

Fuse qknorm+rope+insert (#20)

2026-06-03 23:10:47 +08:00

moe

[Perf] Optimize moe permute by pre-allocate buffer, 9~14% kernel performance improvement (#43014)

2026-05-28 06:18:26 -07:00

quantization

[9/n] Migrate attention and cache kernels to torch stable ABI (continued) (#43717)

2026-05-29 04:44:45 +00:00

quickreduce

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

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

rocm

[Kernel][ROCm] Native W4A16 kernel for AMD RDNA3 (gfx1100) — fp16 + bf16 (#41394)

2026-05-29 11:04:40 +00:00

async_util.cuh

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

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

cache.h

Use CU_MEMCPY_SRC_ACCESS_ORDER_ANY for batch KV cache swaps (#39306)

2026-05-10 05:57:09 +03:00

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_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

fix(rocm): remove workaround causing invalid argument on Qwen3.5 with TP=2 (#40686)

2026-05-06 01:38:32 -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

fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

[Feat][DSV4] Fuse q pad into deepseek v4 fused kernel (#43162)

2026-05-26 05:12:54 -10:00

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

minimax_m3_build_k2q_csr.cu

Prefill sparse attention CuteDSL kernel (#10)

2026-06-01 06:19:03 +08:00

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

ops.h

Prefill sparse attention CuteDSL kernel (#10)

2026-06-01 06:19:03 +08:00

persistent_topk.cuh

[8/n] Migrate merge_attn_states, mamba, sampler to torch stable ABI (continued) (#43361)

2026-05-27 09:35:24 -07:00

spinloop.cpp

Add VLLM_USE_SPINLOOP_EXT to use more efficient busy polling (#36517)

2026-05-11 16:11:49 -07:00

torch_bindings.cpp

Prefill sparse attention CuteDSL kernel (#10)

2026-06-01 06:19:03 +08:00

torch_utils.h

[6/n] Migrate activation kernels, gptq, gguf, non cutlass w8a8 to libtorch stable ABI (continued) (#42663)

2026-05-20 00:18:12 -07:00

type_convert.cuh

[7/n] Migrate pos_encoding and norm kernels to libtorch stable ABI (continued) (#43209)

2026-05-23 13:20:00 +08:00