feat(unsloth): support Nemotron-H hybrid MoE fine-tuning on DGX Spark#338
Merged
Conversation
albcui
marked this pull request as ready for review
June 15, 2026 20:09
albcui
changed the title
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Enterprise Run ID: 📒 Files selected for processing (9)
✅ Files skipped from review due to trivial changes (1)
🚧 Files skipped from review as they are similar to previous changes (8)
📝 WalkthroughWalkthroughAdds prebuilt CUDA extension wheels ( ChangesUnsloth training: prebuilt CUDA wheels + device_map support
Possibly Related PRs
Suggested Reviewers
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (2)
docker/Dockerfile.nmp-unsloth-training (1)
130-134: ⚡ Quick winWheel path version inconsistency.
mamba-ssmpath hardcodes version2.3.0whilecausal-conv1duses a full wildcard. If multiplecausal-conv1dversions exist in the wheel directory, the wildcard could match unexpectedly. Consider hardcoding the causal-conv1d version for consistency (e.g.,causal_conv1d-1.5.3-cp312*.whl), or document why version flexibility is intentional here.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@docker/Dockerfile.nmp-unsloth-training` around lines 130 - 134, The causal-conv1d wheel path uses a version wildcard pattern while mamba-ssm hardcodes its version, creating inconsistency that could cause unpredictable wheel selection if multiple versions exist. Modify the causal_conv1d wheel path in the RUN command (currently using `causal_conv1d-*cp312*.whl`) to hardcode a specific version number (e.g., `causal_conv1d-1.5.3-cp312*.whl`) to match the explicit versioning approach used for mamba_ssm-2.3.0-cp312*.whl, ensuring consistent and predictable dependency installation.services/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.py (1)
22-39: ⚡ Quick winCross-service inconsistency: automodel still raises.
The automodel service's
get_percentage(atservices/automodel/src/nmp/automodel/tasks/file_io/callbacks.py:18-45) still raisesValueErroron invalid inputs. This divergence could confuse maintainers when similar functions behave differently across services.Consider aligning both implementations to use the clamping approach.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@services/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.py` around lines 22 - 39, The automodel service's get_percentage function still raises ValueError on invalid inputs, while the unsloth service's get_percentage function now uses the clamping approach to handle out-of-range values gracefully. Align the automodel service's get_percentage function to match the unsloth implementation by replacing the ValueError raising behavior with the same clamping logic (using max/min to constrain current between 0 and total, and returning 0 when total is invalid), while also adding the debug logging that explains why clamping occurs. This ensures consistent error handling across services and prevents maintainer confusion.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@packages/nmp_platform_runner/src/nmp/platform_runner/config/local.yaml`:
- Around line 32-36: The CPU executor configuration in the local.yaml file has a
mismatch between its provider and profile settings. The `provider: cpu` at line
32 is paired with `profile: gpu` at line 33, which is incorrect. Change the
profile value from gpu to cpu to ensure the configuration properly reflects CPU
resources and jobs are scheduled with the correct resource expectations.
- Line 36: The launcher_tool_path in local.yaml uses a relative path that
depends on the repository being available at runtime and the platform starting
from the repo root. Replace the relative path
`./services/core/jobs/jobs-launcher/jobs-launcher` with an absolute path
pointing to the built jobs-launcher binary (e.g.,
`/full/path/to/nemo-platform/services/core/jobs/jobs-launcher/jobs-launcher`),
and ensure the documentation or configuration comments clarify that users must
first run `make build-jobs-launcher` to build the Go binary before the platform
can properly inject the launcher into containers.
---
Nitpick comments:
In `@docker/Dockerfile.nmp-unsloth-training`:
- Around line 130-134: The causal-conv1d wheel path uses a version wildcard
pattern while mamba-ssm hardcodes its version, creating inconsistency that could
cause unpredictable wheel selection if multiple versions exist. Modify the
causal_conv1d wheel path in the RUN command (currently using
`causal_conv1d-*cp312*.whl`) to hardcode a specific version number (e.g.,
`causal_conv1d-1.5.3-cp312*.whl`) to match the explicit versioning approach used
for mamba_ssm-2.3.0-cp312*.whl, ensuring consistent and predictable dependency
installation.
In `@services/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.py`:
- Around line 22-39: The automodel service's get_percentage function still
raises ValueError on invalid inputs, while the unsloth service's get_percentage
function now uses the clamping approach to handle out-of-range values
gracefully. Align the automodel service's get_percentage function to match the
unsloth implementation by replacing the ValueError raising behavior with the
same clamping logic (using max/min to constrain current between 0 and total, and
returning 0 when total is invalid), while also adding the debug logging that
explains why clamping occurs. This ensures consistent error handling across
services and prevents maintainer confusion.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: d32a186a-53d9-428d-8265-8e81d5aa8312
📒 Files selected for processing (8)
docker-bake.hcldocker/Dockerfile.nmp-unsloth-trainingdocker/unsloth/README.mdpackages/nmp_platform_runner/src/nmp/platform_runner/config/local.yamlplugins/nemo-customizer/src/nemo_customizer/skills/nemo-customizer/references/hyperparameters.mdservices/unsloth/src/nmp/unsloth/schemas.pyservices/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.pyservices/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.py
Contributor
|
albcui
force-pushed
the
albcui/customizer/unsloth-nemotron
branch
2 times, most recently
from
June 15, 2026 20:39
a522cdd to
9146ab2
Compare
albcui
changed the title
albcui
force-pushed
the
albcui/customizer/unsloth-nemotron
branch
from
June 16, 2026 18:13
9146ab2 to
926b0f0
Compare
Contributor
There was a problem hiding this comment.
🧹 Nitpick comments (1)
packages/nmp_platform/config/local.yaml (1)
78-87: ⚡ Quick winRemove redundant
launcher_tool_pathfrom executor entries.Both executor entries specify
launcher_tool_path(lines 82, 87) with the same value already defined inexecutor_defaults.docker(line 95). The executor-level config unnecessarily duplicates the default.♻️ Remove redundant configuration
- provider: cpu profile: gpu backend: docker - config: - launcher_tool_path: ./services/core/jobs/jobs-launcher/jobs-launcher - provider: gpu profile: gpu backend: docker - config: - launcher_tool_path: ./services/core/jobs/jobs-launcher/jobs-launcher🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/nmp_platform/config/local.yaml` around lines 78 - 87, Remove the redundant launcher_tool_path configuration from both executor entries (the cpu provider and gpu provider executor blocks) in the config file. The launcher_tool_path value is already defined once in the executor_defaults.docker section, so specifying it individually in each executor's config block creates unnecessary duplication. Delete the launcher_tool_path lines from both the cpu executor entry and the gpu executor entry, keeping only the entries in the shared defaults section.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Nitpick comments:
In `@packages/nmp_platform/config/local.yaml`:
- Around line 78-87: Remove the redundant launcher_tool_path configuration from
both executor entries (the cpu provider and gpu provider executor blocks) in the
config file. The launcher_tool_path value is already defined once in the
executor_defaults.docker section, so specifying it individually in each
executor's config block creates unnecessary duplication. Delete the
launcher_tool_path lines from both the cpu executor entry and the gpu executor
entry, keeping only the entries in the shared defaults section.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 7e3a151c-1e29-4c39-8809-a9a988683e7a
📒 Files selected for processing (9)
docker-bake.hcldocker/Dockerfile.nmp-unsloth-trainingdocker/unsloth/README.mdpackages/nmp_platform/config/local.yamlpackages/nmp_platform_runner/src/nmp/platform_runner/config/local.yamlplugins/nemo-customizer/src/nemo_customizer/skills/nemo-customizer/references/hyperparameters.mdservices/unsloth/src/nmp/unsloth/schemas.pyservices/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.pyservices/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.py
🚧 Files skipped from review as they are similar to previous changes (7)
- services/unsloth/src/nmp/unsloth/tasks/training/backends/unsloth_sft.py
- plugins/nemo-customizer/src/nemo_customizer/skills/nemo-customizer/references/hyperparameters.md
- services/unsloth/src/nmp/unsloth/schemas.py
- docker/unsloth/README.md
- docker-bake.hcl
- services/unsloth/src/nmp/unsloth/tasks/file_io/callbacks.py
- docker/Dockerfile.nmp-unsloth-training
soluwalana
approved these changes
Jun 16, 2026
anubhutivyas
approved these changes
Jun 16, 2026
github-merge-queue
Bot
removed this pull request from the merge queue due to a conflict with the base branch
Jun 16, 2026
Enable fine-tuning of hybrid Mamba-2 + MoE models (NVIDIA Nemotron-H,
e.g. Nemotron-3-Nano-30B-A3B) on the Unsloth customization backend, plus
two supporting fixes uncovered bringing the backend up on a unified-memory
GB10 (DGX Spark, arm64/Blackwell).
- Add mamba-ssm + causal-conv1d to the nmp-unsloth-training image, installed
from the same prebuilt cu13.1.1/cp312 wheels as nmp-automodel-base via the
causal-conv1d-wheel / mamba-ssm-wheel bake contexts (versions single-sourced,
no recompile). Nemotron-H remote code imports mamba_ssm at load time.
- Add configurable model.device_map (default {"": 0}). Without an explicit
single-device map, accelerate's device_map="auto" sizes the GPU budget from
CUDA free memory, which on unified-memory parts reports a small dynamic
carve-out and spills layers to CPU -- aborting 4-bit loads.
- Fix file_io download crash: get_percentage now clamps instead of raising when
the live file count exceeds the pre-listing total (off-by-one on sources with
nested directories), so a cosmetic progress value can't fail a multi-GB
download mid-transfer.
- Register cpu/gpu + gpu/gpu docker execution profiles in the local runner
config so the customization backends run out-of-the-box on a local platform.
- Docs: document model.device_map and the 16-bit-LoRA path for hybrid MoE models
in the customizer hyperparameters reference; refresh the unsloth image README
build steps (shared wheels / USE_LOCAL_WHEELS, accurate step list).
Validated end-to-end on DGX Spark: Nemotron-3-Nano-30B-A3B loads and trains in
16-bit LoRA (load_in_4bit=false).
Signed-off-by: Albert Cui <[email protected]>
Signed-off-by: Albert Cui <[email protected]>
albcui
force-pushed
the
albcui/customizer/unsloth-nemotron
branch
from
June 16, 2026 20:52
926b0f0 to
1cae724
Compare
albcui
enabled auto-merge
June 16, 2026 20:52
This was referenced Jun 25, 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
3 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.
Enable fine-tuning of hybrid Mamba-2 + MoE models (NVIDIA Nemotron-H, e.g. Nemotron-3-Nano-30B-A3B) on the Unsloth customization backend, plus two supporting fixes uncovered bringing the backend up on a unified-memory GB10 (DGX Spark, arm64/Blackwell).
Validated end-to-end on DGX Spark: Nemotron-3-Nano-30B-A3B loads and trains in 16-bit LoRA (load_in_4bit=false).
Summary by CodeRabbit
device_mapsupport for Unsloth model loading to control GPU placement (including single-GPU defaults and explicit mappings).