[Model] Add DeepSeek-V4 CPU backend (#55355)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2026-09-10 14:14:48 +08:00
committed by GitHub
parent 49632fb3ac
commit c3ccc0e957
47 changed files with 9564 additions and 627 deletions
+46 -4
View File
@@ -203,14 +203,26 @@ class CpuPlatform(Platform):
cache_config = vllm_config.cache_config
is_deepseek_v4 = (
model_config is not None
and getattr(model_config.hf_config, "model_type", None) == "deepseek_v4"
)
# The CPU MLA decode kernel only compiles with block_size=16 today
# (see csrc/cpu/mla_decode.cpp). If the model uses MLA we override
# the default block size regardless of user preference to avoid a
# runtime kernel dispatch failure. AMX MLA has no such constraint
# (same AMX-available condition as get_attn_backend_cls), so it's
# excluded from this override.
cpu_mla_enabled = model_config is not None and getattr(
model_config, "use_mla", False
# excluded from this override. DeepSeek-V4 is also excluded here and
# handled in its own branch below: its sparse-MLA and indexer
# backends declare block_size=256 as their only supported kernel
# block size (DeepseekV4SparseMLABackend/DeepseekV4IndexerBackend),
# so it needs the same override-regardless-of-preference treatment
# as CPU MLA, just with a different value.
cpu_mla_enabled = (
not is_deepseek_v4
and model_config is not None
and getattr(model_config, "use_mla", False)
)
amx_mla_enabled = (
cpu_mla_enabled
@@ -219,6 +231,21 @@ class CpuPlatform(Platform):
and vllm_config.attention_config.backend != AttentionBackendEnum.CPU_MLA
)
reference_cpu_mla_enabled = cpu_mla_enabled and not amx_mla_enabled
# DeepSeek-V4's CPU attention/indexer kernels
# (csrc/cpu/sgl-kernels/{flash_mla,store_cache,compressor,
# paged_mqa_logits,topk}.cpp) are AMX-kernel-backed and built on the
# same paged/position-indexed conventions as the GPU/XPU backends
# (block_table/slot_mapping addressing throughout, chunk metadata
# that already carries per-token causal offsets for partial/extend
# continuation) -- so they support chunked prefill and prefix
# caching the same way. `amx_mla_enabled` above deliberately
# excludes DeepSeek-V4 (it has its own block-size requirements,
# unrelated to chunked-prefill support), so it can't be reused here.
amx_mla_or_dsv4_enabled = amx_mla_enabled or (
is_deepseek_v4
and cls.get_cpu_architecture() == CpuArchEnum.X86
and torch.cpu._is_amx_tile_supported()
)
if reference_cpu_mla_enabled:
if cache_config.user_specified_block_size and cache_config.block_size != 16:
logger.warning(
@@ -227,6 +254,17 @@ class CpuPlatform(Platform):
cache_config.block_size,
)
cache_config.block_size = 16
elif is_deepseek_v4:
if (
cache_config.user_specified_block_size
and cache_config.block_size != 256
):
logger.warning(
"DeepSeek-V4 CPU backend requires block_size=256, "
"overriding user-specified block_size=%s.",
cache_config.block_size,
)
cache_config.block_size = 256
elif not cache_config.user_specified_block_size:
cache_config.block_size = 128
@@ -439,7 +477,11 @@ class CpuPlatform(Platform):
vllm_config.parallel_config.tensor_parallel_size
)
if model_config is not None and model_config.use_mla and not amx_mla_enabled:
if (
model_config is not None
and model_config.use_mla
and not amx_mla_or_dsv4_enabled
):
logger.info_once(
"MLA is enabled on a non-GPU platform; forcing chunked "
"prefill and prefix caching to be disabled."