[New Model] DeepSeek-V3.2 (Rebased to Main) (#25896)
Signed-off-by: Chen Zhang <zhangch99@outlook.com> Signed-off-by: youkaichao <youkaichao@gmail.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Signed-off-by: mgoin <mgoin64@gmail.com> Signed-off-by: NickLucche <nlucches@redhat.com> Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Signed-off-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com> Signed-off-by: Lucia Fang <fanglu@meta.com> Co-authored-by: Chen Zhang <zhangch99@outlook.com> Co-authored-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com> Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com> Co-authored-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: mgoin <mgoin64@gmail.com> Co-authored-by: Lucia Fang <116399278+luccafong@users.noreply.github.com> Co-authored-by: Lucia Fang <fanglu@meta.com> Co-authored-by: NickLucche <nlucches@redhat.com> Co-authored-by: Siyuan Fu <siyuanf@nvidia.com> Co-authored-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Xiaozhu Meng <mxz297@gmail.com> Co-authored-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
This commit is contained in:
co-authored by
Chen Zhang
youkaichao
Lucas Wilkinson
Robert Shaw
Lucas Wilkinson
yewentao256
Wentao Ye
mgoin
Lucia Fang
Lucia Fang
NickLucche
Siyuan Fu
Matthew Bonanni
Xiaozhu Meng
Barry Kang
parent
e23cacda35
commit
fa7e254a7f
@@ -346,8 +346,7 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
|
||||
block_size=1,
|
||||
num_kv_heads=model_config.get_num_kv_heads(parallel_config),
|
||||
head_size=model_config.get_head_size(),
|
||||
dtype=kv_cache_dtype,
|
||||
use_mla=model_config.use_mla).page_size_bytes
|
||||
dtype=kv_cache_dtype).page_size_bytes
|
||||
|
||||
model_cls, _ = ModelRegistry.resolve_model_cls(
|
||||
model_config.architecture,
|
||||
@@ -401,6 +400,31 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
|
||||
"exactly equal.", mamba_padding_pct)
|
||||
|
||||
|
||||
class DeepseekV3ForCausalLM(VerifyAndUpdateConfig):
|
||||
|
||||
@classmethod
|
||||
def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
|
||||
"""
|
||||
Updated fp8 cache to custom "fp8_ds_mla" format for DeepSeekV32
|
||||
"""
|
||||
hf_config = vllm_config.model_config.hf_config
|
||||
|
||||
is_v32 = hasattr(hf_config, "index_topk")
|
||||
|
||||
if is_v32:
|
||||
# For DeepSeekV3.2, we use a custom fp8 format as default (i.e.
|
||||
# "auto")
|
||||
cache_config = vllm_config.cache_config
|
||||
if cache_config.cache_dtype == "auto" or \
|
||||
cache_config.cache_dtype.startswith("fp8"):
|
||||
cache_config.cache_dtype = "fp8_ds_mla"
|
||||
logger.info(
|
||||
"Using custom fp8 kv-cache format for DeepSeekV3.2")
|
||||
if cache_config.cache_dtype == "bfloat16":
|
||||
cache_config.cache_dtype = "auto"
|
||||
logger.info("Using bfloat16 kv-cache for DeepSeekV3.2")
|
||||
|
||||
|
||||
MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
|
||||
"GteModel": SnowflakeGteNewModelConfig,
|
||||
"GteNewModel": GteNewModelConfig,
|
||||
@@ -417,4 +441,5 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
|
||||
"MambaForCausalLM": MambaModelConfig,
|
||||
"Mamba2ForCausalLM": MambaModelConfig,
|
||||
"FalconMambaForCausalLM": MambaModelConfig,
|
||||
"DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user