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

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Misha Goin
2026-10-12 09:46:22 -07:00
committed by GitHub
co-authored by Claude
parent e0317b28ce
commit d79b52b182
242 changed files with 2783 additions and 3638 deletions
+2 -1
View File
@@ -68,7 +68,8 @@ class FusedQKVRMSNormKernel(VllmTritonJitKernel["FusedQKVRMSNormKernel.CompileKe
weight_ptr = kv_weight_ptr
row_out = kv_out_ptr + token_idx * kv_out_stride
# RMSNorm in fp32 throughout — matches csrc/layernorm_kernels.cu's
# RMSNorm in fp32 throughout — matches
# csrc/layernorm_kernels.cu's
# `(scalar_t)(x * s_variance * w)` and DeepseekV4's compressor kernel, which
# keep x, rrms, and w all in fp32 and perform a single cast at store.
block = tl.arange(0, BLOCK_SIZE)
+1 -1
View File
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""ROCm entry point for the fused Kimi-K3 KDA chunk kernel.
The kernel in ``csrc/libtorch_stable/kimi_k3/fused_kda_chunk_kernel_rocm.cu``
The kernel in ``csrc/kimi_k3/fused_kda_chunk_kernel_rocm.cu``
replaces the chunk-state recurrence and the output GEMM of the Triton chunk
path with a single launch that keeps the per-chunk state in registers, so the
``[chunks, H, V, K]`` state tensor and the recomputed values never reach HBM.
+1 -1
View File
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""ROCm entry points for the fused Kimi-K3 KDA decode kernel.
The kernel in ``csrc/libtorch_stable/kimi_k3/fused_kda_decode_kernel_rocm.cu``
The kernel in ``csrc/kimi_k3/fused_kda_decode_kernel_rocm.cu``
replaces, for a pure non-speculative decode batch, the three Triton launches
and two copies the AMD KDA layer otherwise runs per layer: the packed causal
conv1d update, the recurrent delta-rule step, and the gated output RMSNorm.
@@ -3,7 +3,7 @@
"""Fused MLA prefill and decode epilogues for Kimi-K3.
Thin wrappers over the CUDA ops in
``csrc/libtorch_stable/fused_kimi_k3_mla_key_concat_kv_cache_kernel.cu``, which
``csrc/fused_kimi_k3_mla_key_concat_kv_cache_kernel.cu``, which
mirror ``fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_{bf16,fp8}_insert``.
- ``fused_mla_key_concat_kv_cache_insert`` (bf16): optionally apply RoPE,