Skip to content

Fix MTP with Deepseek R1 Fp4 #7376

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 9 commits into from
Jun 24, 2025
Merged

Fix MTP with Deepseek R1 Fp4 #7376

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
65490a2
Fix MTP with Deepseek R1 Fp4
pyc96 Jun 20, 2025
c653b1b
Merge branch 'main' into mtp-fix
zhyncs Jun 20, 2025
bf8af63
Merge branch 'main' into mtp-fix
pyc96 Jun 20, 2025
241d55f
Fix comment
pyc96 Jun 21, 2025
9d6da0f
Merge branch 'main' into mtp-fix
zhyncs Jun 22, 2025
357bbf8
fix
pyc96 Jun 22, 2025
1baaaea
Merge branch 'main' into mtp-fix
zhyncs Jun 23, 2025
6c2c8a7
Merge branch 'main' into mtp-fix
zhyncs Jun 23, 2025
055ee31
Merge branch 'main' into mtp-fix
zhyncs Jun 24, 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
6 changes: 6 additions & 0 deletions python/sglang/srt/layers/moe/fused_moe_triton/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,6 +330,12 @@ def __init__(
self.tp_rank = get_tensor_model_parallel_rank()
self.num_experts = num_experts
self.expert_map = None

if enable_flashinfer_moe and quant_config is None:
logger.warning("Disable flashinfer MoE when quantization config is None.")
enable_flashinfer_moe = False
enable_ep_moe = False

Comment on lines +333 to +338
Copy link
Contributor

@Jonahcb Jonahcb Oct 29, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@pyc96 what was the justification for adding this? Does Flashinfer's Cutluss MoE implementation only support quantized models? At the bottom of this Flashinfer documentation it says that Flashinfer Cutlass Moe backend supports most data types: https://docs.flashinfer.ai/generated/flashinfer.fused_moe.cutlass_fused_moe.html#flashinfer.fused_moe.cutlass_fused_moe

Thanks!

self.enable_flashinfer_moe = enable_flashinfer_moe
if enable_ep_moe:
assert (
Expand Down
6 changes: 6 additions & 0 deletions python/sglang/srt/models/deepseek_nextn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,12 @@ def __init__(
prefix: str = "",
) -> None:
super().__init__()
if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
logger.warning(
"Overriding DeepseekV3ForCausalLMNextN quant config for modelopt_fp4 Deepseek model."
)
quant_config = None

self.vocab_size = config.vocab_size

self.embed_tokens = VocabParallelEmbedding(
Expand Down
9 changes: 8 additions & 1 deletion python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2201,7 +2201,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=Fal
q_a_proj_weight = cached_a_proj[q_a_proj_name]
kv_a_proj_weight = cached_a_proj[kv_a_proj_name]
cat_dim = 0
if (
if self.quant_config is not None and (
self.quant_config.get_name() == "awq"
or self.quant_config.get_name() == "moe_wna16"
):
Comment on lines +2204 to 2207
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Good addition of the self.quant_config is not None check. This prevents potential AttributeError if self.quant_config happens to be None, which is now possible due to the changes in python/sglang/srt/model_loader/loader.py for the Deepseek R1 Fp4 model.

Expand Down Expand Up @@ -2232,6 +2232,13 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=Fal
for scale in ["k_scale", "v_scale"]:
if scale in name:
name = name.replace(f"{scale[0]}_proj", "attn_mqa")
break
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Adding break here is a good optimization. It ensures the loop terminates as soon as the scale is found and name is updated, avoiding unnecessary iterations.

if name not in params_dict:
# modelopt ckpt contains not needed weights for MTP module:
# model.decoder.self_attn.attn_mqa.v_scale and
# model.decoder.self_attn.attn_mqa.k_scale
logger.warning(f"{name} not found in params_dict.")
continue
Comment on lines +2236 to +2241
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

This check to ensure name is in params_dict before attempting to access it is a robust addition. The warning message clearly explains why certain weights might be skipped, which is helpful for debugging and understanding model loading behavior, especially with modelopt checkpoints that might contain extra, unneeded weights for specific modules like MTP.

param = params_dict[name]
weight_loader = getattr(
param, "weight_loader", default_weight_loader
Expand Down
Loading