Repository navigation
Shard MuseSpark's experts whole under tensor parallelism - #49787
Closed
IlyasMoutawwakil wants to merge 9 commits into
Closed
IlyasMoutawwakil wants to merge 9 commits into
IlyasMoutawwakil wants to merge 9 commits into
Conversation
should_convert_module matched an entry from the start of the module name with
no boundary after it, so "model.layers.1" also skipped model.layers.10 to 19,
and Qwen3.8-27B-FP8's router entry "mlp.gate" also skipped the quantized
mlp.gate_proj (its FP8 weight then reached a bf16 linear). A match from the
start now has to end on a "." it consumed, or right before a "." or the end
of the name, which keeps regex entries that carry their own boundary such as
"^model\.audio_tower(?:\.|$)". The endswith clause Hub skip-lists rely on
("gate" covering shared_expert_gate) is unchanged.
Supersedes #47990, which anchors with "$" and so stops matching such regex
entries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… file Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
process_target_pattern unescaped \. before extracting the capturing group, so the reversed rename's group matched any character in place of a dot. Unescape only outside the group. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ision mlp keys The `.block_sparse_moe.` -> `.mlp.` rename reversed on save turned every `.mlp.` key back into `.block_sparse_moe.`, the dense and vision ones too. Name the MoE block's children in the pattern; the reversed group keeps its escaped dots. Run the reverse-mapping test that was skipped, on the composite model only, as for Llava. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MuseSpark normalizes each expert's output before the ranks combine them, so an expert split across ranks cannot compute its norm. Its TP plan now gives the MoE the expert-parallel layout (router masking, whole experts per rank, the norm replicated with its gradient all-reduced), and its EP plan replicates the router. The model test drops its override of test_moe_parallel_plans_shard_experts: the common test now passes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
IlyasMoutawwakil
force-pushed
the
fg-split/fix-muse-spark-moe-plan
branch
from
October 20, 2026 16:04
ac9c920 to
72c152d
Compare
IlyasMoutawwakil
force-pushed
the
fg-split/fix-minimax-m3-vl-save-mapping
branch
from
October 20, 2026 16:04
daad369 to
5dd6ef5
Compare
vasqu
reviewed
Oct 20, 2026
vasqu
left a comment
vasqu
left a comment
Collaborator
There was a problem hiding this comment.
Imo this doesn't need to be fixed, don't feel like the TP here for moe is important
| "layers.*.mlp.experts": "moe_tp_experts", | ||
| } | ||
| base_model_ep_plan = { | ||
| "layers.*.mlp.gate": "replicate", |
Collaborator
There was a problem hiding this comment.
? why is this suddenly needed?
Comment on lines
+113
to
+118
| # Post-expert normalization requires complete expert outputs before ranks combine them. | ||
| "layers.*.mlp.gate": "ep_router", | ||
| "layers.*.mlp.experts.gate_up_proj": "grouped_gemm", | ||
| "layers.*.mlp.experts.down_proj": "grouped_gemm", | ||
| "layers.*.mlp.experts.post_expert_norm": "replicated_with_grad_allreduce", | ||
| "layers.*.mlp.experts": "moe_tp_experts", |
Collaborator
There was a problem hiding this comment.
do we really need this? Werent we always using colwise/rowwise for TP. Kinda feels like a conflict between what we keep under TP and what under EP (esp after the split). Otherwise we would need to update all moe tps no?
Member
Author
There was a problem hiding this comment.
agree, it's an ugly workaround
Member
Author
|
yeh it's not worth it, will just skip tp quantization checks for it |
IlyasMoutawwakil
force-pushed
the
fg-split/fix-minimax-m3-vl-save-mapping
branch
5 times, most recently
from
October 23, 2026 13:35
5a70522 to
19d9f5c
Compare
…mapping' into fg-split/fix-muse-spark-moe-plan # Conflicts: # src/transformers/conversion_mapping.py # src/transformers/quantizers/quantizers_utils.py # tests/quantization/test_quantizers_utils.py # tests/utils/test_core_model_loading.py
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: muse_spark |
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
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.
What. MuseSpark normalizes each expert's output before the ranks combine it, so an expert split across ranks
cannot compute its norm. The TP plan now gives the MoE the expert-parallel layout: router masking, whole experts per
rank, and the norm replicated with an all-reduced gradient. The EP plan replicates the router. The model test drops
its override of
test_moe_parallel_plans_shard_experts, since the common test now passes.Depends on. the MiniMax-M3-VL fix branch (#49786, stacked only).
How tested.
tests/models/muse_spark.🤖 Generated with Claude Code