[model] fix: register both MTP submodule spellings in qwen3_next_bridge#11
Open
Zhichenzzz wants to merge 1 commit into
Open
[model] fix: register both MTP submodule spellings in qwen3_next_bridge#11Zhichenzzz wants to merge 1 commit into
Zhichenzzz wants to merge 1 commit into
Conversation
This was referenced Jun 10, 2026
…bridge Megatron-LM renamed the MTP submodule transformer_layer -> mtp_model_layer. Register every qwen3_next MTP weight mapping (MoE router, layernorms, attention, experts, shared expert) under both spellings so the bridge converts checkpoints from either Megatron-LM version, mirroring the miles converter change.
Zhichenzzz
force-pushed
the
fix/qwen3next-mtp-rename
branch
from
4728c32 to
a06b06d
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Megatron-LM renamed the MTP submodule
transformer_layer→mtp_model_layer(upstream NVIDIA main has no old-name API surface left).qwen3_next_bridgestill registered only the old spelling, so MTP weights fail to map against a renamed Megatron-LM — whileqwen35_vl_bridgeanddeepseek/common.pyalready expect the new name, andglm45_bridgehandles both.Register both spellings (glm45's approach) for all 12 MTP mappings (6 AutoMapping dict entries + QKV/GatedMLP/Replicated special mappings) so the bridge works with either Megatron-LM version. Mapping entries whose megatron param does not exist at runtime are never consulted, so the extra spelling is inert.
Validated:
mapping_registry().megatron_to_hf_lookupresolves all 8 probe names (both spellings × router/QKV/expert-fc1/shared-gate) to the correct mapping types.Companion PRs: radixark/Megatron-LM#54 (the rename itself), radixark/miles#1307 (miles converters).