Skip to content

dspark: add shared expert weight padding for block-quantized MoE #33

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

Closed
allenh1 wants to merge 1 commit into from
+13 −3
Closed

dspark: add shared expert weight padding for block-quantized MoE #33

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
16 changes: 13 additions & 3 deletions vllm/models/deepseek_v4/nvidia/dspark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -700,21 +700,25 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
self.max_batch = max(1, int(vllm_config.scheduler_config.max_num_seqs))
hidden_size = int(config.hidden_size)
dtype = vllm_config.model_config.dtype
quant_config = get_draft_quant_config(vllm_config)
self.quant_config = get_draft_quant_config(vllm_config)
self.pad_shared_expert = (
getattr(self.quant_config, "weight_block_size", None) is not None
and not vllm_config.parallel_config.use_sequence_parallel_moe
)
self.dspark_aux_hidden_size = hidden_size * len(self.target_layer_ids)

self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
hidden_size,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}embed_tokens",
)
self.main_proj = ReplicatedLinear(
self.dspark_aux_hidden_size,
hidden_size,
bias=False,
params_dtype=dtype,
quant_config=quant_config,
quant_config=self.quant_config,
prefix=f"{prefix}mtp.0.main_proj",
return_bias=False,
)
Expand Down Expand Up @@ -1044,6 +1048,12 @@ def map_name(name: str) -> str | None:
else ".weight_scale_inv"
)
name = name.removesuffix(".scale") + suffix
if ".shared_experts.w2" in name:
name = name.replace(".shared_experts.w2", ".shared_experts.down_proj")
if self.pad_shared_expert and ".shared_experts." in name:
loaded_weight = DeepseekV4Model._pad_shared_expert_weight(
self.quant_config, name, loaded_weight
)
for param_name, weight_name, shard_id in stacked_params_mapping:
if not name.startswith("layers."):
continue
Expand Down