biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Attention] Use sparse prefill kernel for fp8 kv-cache in DeepSeek-v3.2 (#27532)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson
2025-12-12 08:57:47 -05:00
committed by GitHub
parent 91401c7a26
commit 3e41992fec
30 changed files with 1372 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
csrc/torch_bindings.cpp
View File

@@ -754,6 +754,13 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
"Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
cache_ops.impl("cp_gather_cache", torch::kCUDA, &cp_gather_cache);
cache_ops.def(
"cp_gather_and_upconvert_fp8_kv_cache(Tensor src_cache, Tensor! dst, "
"Tensor block_table, Tensor seq_lens, Tensor workspace_starts, int "
"batch_size) -> ()");
cache_ops.impl("cp_gather_and_upconvert_fp8_kv_cache", torch::kCUDA,
&cp_gather_and_upconvert_fp8_kv_cache);
cache_ops.def(
"indexer_k_quant_and_cache(Tensor k, Tensor! kv_cache, Tensor "
"slot_mapping, "