Repository navigation
Match a skip-list entry from the start only up to a module boundary - #49771
Merged
Merged
Conversation
IlyasMoutawwakil
force-pushed
the
fg-split/fix-skip-list-index
branch
from
October 20, 2026 11:08
0489006 to
1d88f2d
Compare
|
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. |
IlyasMoutawwakil
force-pushed
the
fg-split/fix-skip-list-index
branch
from
October 20, 2026 11:20
1d88f2d to
a063a21
Compare
IlyasMoutawwakil
force-pushed
the
fg-split/fix-skip-list-index
branch
from
October 20, 2026 13:33
a063a21 to
552b916
Compare
vasqu
reviewed
Oct 20, 2026
Comment on lines
-27
to
-36
| """Return False if an entry of `patterns` covers `full_name`, True otherwise. | ||
|
|
||
| An entry covers the names it matches from the start up to a module boundary, and the names ending with it. | ||
| So `"model.layers.1"` covers `model.layers.1.mlp` but not `model.layers.10`. | ||
| """ | ||
| if patterns is None: | ||
| return True | ||
|
|
||
| # We should avoid converting in the following situations: | ||
| # 1. The pattern appears as a prefix followed by a dot in `full_name` | ||
| # (e.g., "model.decoder.layer.11." matches "model.decoder.layer.11.attn.weight"). | ||
| # 2. The pattern matches `full_name` exactly or via regex | ||
| # (e.g., "lm_head" matches "lm_head"; "model.decoder.layer.*" matches "model.decoder.layer.11.attn.weight"). | ||
| # 3. `full_name` ends with the pattern | ||
| # (e.g., "fc1" matches "model.decoder.layers.23.fc1"). |
Collaborator
There was a problem hiding this comment.
why a new docstring, shouldnt we just adjust the existing cases instead?
Member
Author
There was a problem hiding this comment.
fair, restored the original comment block and adjusted the cases for the boundary rule, docstring's gone
Collaborator
There was a problem hiding this comment.
why is the test in mxfp4? do we not have a general utils test file?
Member
Author
There was a problem hiding this comment.
weirdly that's the test checking it, and no quantization utils test file, i can create one
Member
Author
There was a problem hiding this comment.
moved it to a new tests/quantization/test_quantizers_utils.py with main's cases + the new ones, out of mxfp4
IlyasMoutawwakil
enabled stacked auto-merge
October 22, 2026 16:04
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>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… file Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
IlyasMoutawwakil
force-pushed
the
fg-split/fix-skip-list-index
branch
from
October 23, 2026 13:03
446dc5d to
18d5d46
Compare
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: mxfp4 |
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.
should_convert_modulematched a skip-list entry with a barere.match, so"model.layers.1"also skippedlayers 10 to 19. A match from the start now ends at a module boundary. An entry still covers its sub-tree, a regex
with its own boundary still works, and a name still covers the names ending in it (
"gate"coversshared_expert_gate).Why we need it. Loading real FP8 checkpoints with the fine-grained FP8 support (#49777) blocked on it.
Qwen3.8-27B-FP8 lists its routers as
mlp.gate, which also skipped the quantizedmlp.gate_proj. Its FP8 weightthen reached a bf16 linear. Every quantizer that takes
modules_to_not_convertuses this function, so main has thesame bug, and any skip-list naming a layer by index hits it too. The key is wrapped in a group so the boundary applies
to every alternative of a key with
|.Supersedes #47990, which anchors with
$and so stops matching regex entries that carry their own boundary, such as^model\.audio_tower(?:\.|$).Behaviour change. A skip-list entry no longer covers longer siblings (
model.layers.1vsmodel.layers.10,mlp.gatevsmlp.gate_proj).Depends on. main. First PR of the fine-grained stack.
How tested.
tests/quantization/test_quantizers_utils.py, a new general test file.test_should_convert_modulemoves there from the MXFP4 suite, with the new cases.
The comments and docstrings that say why each change exists are there for readability. Feel free to ask for any of them to go if it's self-evident.
🤖 Generated with Claude Code