Skip to content

[BugFix] Fix default kv-cache-dtype default for DeepseekV3.2 #25988

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Oct 1, 2025
Merged

[BugFix] Fix default kv-cache-dtype default for DeepseekV3.2 #25988

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4c03e2d
fix defaults
LucasWilkinson Sep 30, 2025
d67bee8
add comment
LucasWilkinson Sep 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 14 additions & 14 deletions vllm/model_executor/models/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,7 +400,7 @@ def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
"exactly equal.", mamba_padding_pct)


class DeepseekV3ForCausalLM(VerifyAndUpdateConfig):
class DeepseekV32ForCausalLM(VerifyAndUpdateConfig):

@classmethod
def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
Expand All @@ -409,20 +409,20 @@ def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
"""
hf_config = vllm_config.model_config.hf_config

# Mirror the check in vllm/model_executor/models/deepseek_v2.py
is_v32 = hasattr(hf_config, "index_topk")
assert is_v32

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")
# 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]] = {
Expand All @@ -441,5 +441,5 @@ def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
"MambaForCausalLM": MambaModelConfig,
"Mamba2ForCausalLM": MambaModelConfig,
"FalconMambaForCausalLM": MambaModelConfig,
"DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
"DeepseekV32ForCausalLM": DeepseekV32ForCausalLM,
}