[speechlm2] Temporarily support native Nemotron-V3 with Automodel 0.4 - #15980
Merged
Conversation
Signed-off-by: Dongji Gao <dongjig@nvidia.com>
blisc
approved these changes
Jul 29, 2026
Collaborator
|
/ok to test 6cfdb0b |
blisc
marked this pull request as ready for review
July 29, 2026 00:10
blisc
enabled auto-merge (squash)
July 29, 2026 00:10
blisc
added
the
r3.0.0
Contributor
|
[🤖]: Hi @DongjiGao 👋, We wanted to let you know that a CICD pipeline for this PR just finished successfully. So it might be time to merge this PR or get some approvals. |
blisc
pushed a commit
that referenced
this pull request
Jul 29, 2026
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 changed
model.model.layersto the legacy Automodel 0.4model.backbone.layersaccess while its parallelizer runs._nemotronh_decoder_blockshelper from Automodel PR failed to inference on single input #2638 is present.Why
The Automodel version pinned by the Speech release predates Automodel PR #2638. Its
NemotronHParallelizationStrategyhard-codesmodel.backbone.layers, while the native Nemotron-V3 model exposes decoder blocks atmodel.model.layers. FSDP2 therefore fails during model parallelization withAttributeErrorbefore training starts.Temporary limitation and removal plan
This is deliberately a narrow release-unblocking shim coupled to the exact layer access used by Automodel 0.4. It is not intended to become a general model adapter and does not add TP, PP, or HSDP support for Nemotron-V3.
The shim contains a
TODO(Dongji)and should be removed after the Speech image updates its Automodel dependency to a build containing PR #2638. Until removal, feature detection makes it a no-op on such a fixed build.Validation
nvcr.io/nvidian/nemo-speech:26.06.rc3, digestsha256:7da43274...AttributeError: 'NemotronHForCausalLM' object has no attribute 'backbone'in the pinned parallelizer.5 passed.Full-size validation on the original 8x H200 PCIe QA system is still required before release sign-off.