Skip to content

Update Qwen3 vLLM layer names to match tpu-inference mappings. #3398

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
copybara-service merged 1 commit into from
Mar 12, 2026
+13 −13
Merged

Update Qwen3 vLLM layer names to match tpu-inference mappings. #3398

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
26 changes: 13 additions & 13 deletions src/maxtext/integration/tunix/weight_mapping/qwen3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,12 +67,12 @@ def to_hf_mapping():
return {
# Token embeddings - shard vocab dimension
"base.token_embedder.embedding": (
"model.embed.embedding",
"model.embed_tokens.weight",
("model", None),
),
# Final layer norm - no sharding needed
"base.decoder.decoder_norm.scale": (
"model.norm.scale",
"model.norm.weight",
(None,),
),
# LM head (logits projection) - shard vocab dimension
Expand All @@ -83,49 +83,49 @@ def to_hf_mapping():
# Layer-specific mappings (scanned -> unscanned)
# MLP components - shard hidden dimensions
"base.decoder.layers.mlp.wi_0.kernel": (
"model.layers.*.mlp.gate_proj.kernel",
"model.layers.*.mlp.gate_proj.weight",
(None, "layer", "model"),
),
"base.decoder.layers.mlp.wi_1.kernel": (
"model.layers.*.mlp.up_proj.kernel",
"model.layers.*.mlp.up_proj.weight",
(None, "layer", "model"),
),
"base.decoder.layers.mlp.wo.kernel": (
"model.layers.*.mlp.down_proj.kernel",
"model.layers.*.mlp.down_proj.weight",
("model", "layer", None),
),
# Layer norms - no sharding needed
"base.decoder.layers.pre_self_attention_layer_norm.scale": (
"model.layers.*.input_layernorm.scale",
"model.layers.*.input_layernorm.weight",
(None, "layer"),
),
"base.decoder.layers.post_self_attention_layer_norm.scale": (
"model.layers.*.post_attention_layernorm.scale",
"model.layers.*.post_attention_layernorm.weight",
(None, "layer"),
),
# Attention components - shard head dimensions
"base.decoder.layers.self_attention.query.kernel": (
"model.layers.*.self_attn.q_proj.kernel",
"model.layers.*.self_attn.q_proj.weight",
(None, "layer", "model", None),
),
"base.decoder.layers.self_attention.key.kernel": (
"model.layers.*.self_attn.k_proj.kernel",
"model.layers.*.self_attn.k_proj.weight",
(None, "layer", "model", None),
),
"base.decoder.layers.self_attention.value.kernel": (
"model.layers.*.self_attn.v_proj.kernel",
"model.layers.*.self_attn.v_proj.weight",
(None, "layer", "model", None),
),
"base.decoder.layers.self_attention.out.kernel": (
"model.layers.*.self_attn.o_proj.kernel",
"model.layers.*.self_attn.o_proj.weight",
("model", "layer", None, None),
),
"base.decoder.layers.self_attention.query_norm.scale": (
"model.layers.*.self_attn.q_norm.scale",
"model.layers.*.self_attn.q_norm.weight",
(None, "layer"),
),
"base.decoder.layers.self_attention.key_norm.scale": (
"model.layers.*.self_attn.k_norm.scale",
"model.layers.*.self_attn.k_norm.weight",
(None, "layer"),
),
}
Loading