Nemotron Nano 3 QAD Launcher Example on OSS Nemotron-Post-Training-V2 data - #2134
Merged
Conversation
Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
Contributor
📝 WalkthroughWalkthroughThe PR adds ChangesMegatron-LM workflows
Estimated code review effort: 4 (Complex) | ~45 minutes Suggested reviewers: Sequence Diagram(s)sequenceDiagram
participant ImportTask
participant QuantizationTask
participant QADTrainingTask
participant ExportTask
ImportTask->>QuantizationTask: provide MCore teacher checkpoint
QuantizationTask->>QADTrainingTask: provide NVFP4 student checkpoint
QADTrainingTask->>ExportTask: provide distilled checkpoint
ExportTask->>ExportTask: write final NVFP4 export
🚥 Pre-merge checks | ✅ 6✅ Passed checks (6 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
ChenhanYu
approved these changes
Aug 10, 2026
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Warning
CodeRabbit couldn't request changes on this pull request because it doesn't have sufficient GitHub permissions.
Please grant CodeRabbit Pull requests: Read and write permission and re-run the review.
Actionable comments posted: 4
🤖 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 `@tools/launcher/common/megatron_lm/train/sft.sh`:
- Line 70: Update the MLM_EXTRA_ARGS handling in the sft.sh argument flow to
preserve positional argument boundaries when values contain whitespace, rather
than serializing them with "$*". Use an argument-preserving interface or a
serialization format that the downstream train.sh and finetune.sh scripts
explicitly parse, and add a regression case covering a whitespace-containing
value.
- Around line 61-63: Update the finetune launcher path around the TRAIN_EXE
assignment and MLM_EXTRA_ARGS handling so arguments retain their original
boundaries when passed to finetune.sh. Avoid flattening positional arguments
with "$*" and ensure the downstream invocation preserves values containing
spaces, or explicitly enforce and document space-free values if that is the
intended contract.
In
`@tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml`:
- Around line 52-54: Update QUANT_CFG in
tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml
at lines 52-54 and 133-135 to use MAMBA_MOE_NVFP4_AGGRESSIVE_CFG consistently
for both QAD and export.
- Around line 106-118: Update the environment block for the QAD training
workflow to define QUANT_CFG using the same corrected NVFP4 configuration as the
quantize and export stages, specifically MAMBA_MOE_NVFP4_AGGRESSIVE_CFG. Keep
the existing MLM_MODEL_CFG and other training settings unchanged.
🪄 Autofix
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: c1a01ccc-3521-4678-8bfc-6c8f6e5a6821
📒 Files selected for processing (3)
tools/launcher/common/megatron_lm/train/sft.shtools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yamltools/launcher/modules/Megatron-LM
Comment on lines
+61
to
+63
| finetune) | ||
| TRAIN_EXE=(bash modules/Megatron-LM/examples/post_training/modelopt/finetune.sh) | ||
| ;; |
Contributor
There was a problem hiding this comment.
🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
for script in \
tools/launcher/modules/Megatron-LM/examples/post_training/modelopt/train.sh \
tools/launcher/modules/Megatron-LM/examples/post_training/modelopt/finetune.sh
do
test -f "$script"
printf '\n== %s ==\n' "$script"
rg -n -C 6 \
'MLM_MODEL_CFG|MLM_MODEL_CKPT|MLM_MODEL_SAVE|HF_MODEL_CKPT|MLM_EXTRA_ARGS|\$1|\$@' \
"$script" || true
doneRepository: NVIDIA/Model-Optimizer
Length of output: 160
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== candidate files =='
fd -i -t f 'train\.sh|finetune\.sh' . | rg 'Megatron-LM|modelopt|post_training' || true
printf '%s\n' '== referenced entry points =='
rg -n -C 8 'TRAIN_EXE|finetune\.sh|train\.sh|MLM_MODEL_CFG|MLM_MODEL_CKPT|MLM_MODEL_SAVE|HF_MODEL_CKPT|MLM_EXTRA_ARGS' \
tools/launcher/common/megatron_lm/train/sft.shRepository: NVIDIA/Model-Optimizer
Length of output: 2930
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== tracked Megatron-LM paths =='
git ls-files | rg -i '(^|/)(megatron[-_]lm|modelopt|post_training)(/|$)|finetune\.sh|train\.sh' | head -200 || true
printf '%s\n' '== submodule and launcher context =='
find . -maxdepth 4 -type f \( -name '.gitmodules' -o -name 'README*' -o -name '*.yaml' -o -name '*.yml' \) -print0 |
xargs -0 rg -n -C 3 'Megatron-LM|post_training/modelopt|finetune\.sh|MLM_TRAIN_SCRIPT' 2>/dev/null || true
printf '%s\n' '== repository metadata =='
git remote -v || true
git status --shortRepository: NVIDIA/Model-Optimizer
Length of output: 23135
🏁 Script executed:
#!/bin/bash
set -euo pipefail
base='https://raw.githubusercontent.com/NVIDIA/Megatron-LM/main/examples/post_training/modelopt'
for name in train.sh finetune.sh; do
url="$base/$name"
printf '\n== %s ==\n' "$url"
if curl -fsSL "$url" | rg -n -C 8 \
'MLM_MODEL_CFG|MLM_MODEL_CKPT|MLM_MODEL_SAVE|HF_MODEL_CKPT|MLM_EXTRA_ARGS|\$1|\$@|exec|python|torchrun'; then
:
else
printf '%s\n' 'No matching contract lines or file unavailable'
fi
doneRepository: NVIDIA/Model-Optimizer
Length of output: 2734
🏁 Script executed:
#!/bin/bash
set -euo pipefail
base='https://raw.githubusercontent.com/NVIDIA/Megatron-LM/main/examples/post_training/modelopt'
for name in train.sh finetune.sh; do
url="$base/$name"
printf '\n== %s ==\n' "$url"
curl -fsSL "$url" | sed -n '1,115p'
doneRepository: NVIDIA/Model-Optimizer
Length of output: 5450
🏁 Script executed:
#!/bin/bash
set -euo pipefail
url='https://raw.githubusercontent.com/NVIDIA/Megatron-LM/main/examples/post_training/modelopt/conf/arguments.sh'
printf '== %s ==\n' "$url"
curl -fsSL "$url" | sed -n '1,240p'Repository: NVIDIA/Model-Optimizer
Length of output: 3070
🏁 Script executed:
#!/bin/bash
set -euo pipefail
tree_json=$(curl -fsSL 'https://api.github.com/repos/NVIDIA/Megatron-LM/git/trees/main?recursive=1')
printf '%s\n' '== modelopt configuration files =='
printf '%s' "$tree_json" |
jq -r '.tree[] | select(.path | startswith("examples/post_training/modelopt/conf/")) | .path' |
head -100
printf '%s\n' '== HF_MODEL_CKPT references in modelopt scripts/configuration =='
paths=$(printf '%s' "$tree_json" |
jq -r '.tree[] | select(.type == "blob" and (.path | startswith("examples/post_training/modelopt/"))) | .path')
while IFS= read -r path; do
content=$(curl -fsSL "https://raw.githubusercontent.com/NVIDIA/Megatron-LM/main/$path")
if printf '%s' "$content" | rg -n -C 3 'HF_MODEL_CKPT|TOKENIZER_MODEL|MLM_MODEL_CFG|MLM_MODEL_CKPT|MLM_MODEL_SAVE|MLM_EXTRA_ARGS'; then
printf '== %s ==\n' "$path"
fi
done <<< "$paths"Repository: NVIDIA/Model-Optimizer
Length of output: 33298
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== Megatron-LM submodule entry =='
git ls-files -s -- tools/launcher/modules/Megatron-LM
git config -f .gitmodules --get-regexp 'submodule\..*\.(path|url)' || true
printf '%s\n' '== exact upstream entry-point contracts =='
for name in train.sh finetune.sh; do
printf '\n-- %s --\n' "$name"
curl -fsSL "https://raw.githubusercontent.com/NVIDIA/Megatron-LM/main/examples/post_training/modelopt/$name" |
rg -n 'source .*arguments\.sh|MLM_MODEL_CKPT|MLM_MODEL_SAVE|TOKENIZER_MODEL|MLM_EXTRA_ARGS|MLM_DATA_ARGS|^\$\{LAUNCH_SCRIPT\}|^\$\{LAUNCH_SCRIPT\}.*finetune\.py|^\$\{LAUNCH_SCRIPT\}.*pretrain'
done
printf '%s\n' '== argument-boundary probe for MLM_EXTRA_ARGS="$*" =='
bash -c '
set -- --dataset "path with spaces" --flag
extra="$*"
printf "joined=%q\n" "$extra"
printf "re-expanded words:"
for word in $extra; do printf " <%s>" "$word"; done
printf "\n"
'Repository: NVIDIA/Model-Optimizer
Length of output: 1744
Preserve argument boundaries in MLM_EXTRA_ARGS.
finetune.sh accepts $1 as MLM_MODEL_CFG and consumes the documented environment variables. However, export MLM_EXTRA_ARGS="$*" flattens launcher arguments. Values containing spaces are split by the downstream unquoted expansion. Preserve argument boundaries or restrict and document space-free values.
🤖 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 `@tools/launcher/common/megatron_lm/train/sft.sh` around lines 61 - 63, Update
the finetune launcher path around the TRAIN_EXE assignment and MLM_EXTRA_ARGS
handling so arguments retain their original boundaries when passed to
finetune.sh. Avoid flattening positional arguments with "$*" and ensure the
downstream invocation preserves values containing spaces, or explicitly enforce
and document space-free values if that is the intended contract.
| export MLM_EXTRA_ARGS=${@} | ||
| echo "=== QAD/SFT training ${MLM_MODEL_CFG} (load ${MLM_MODEL_CKPT}) ===" | ||
| ${TRAIN_EXE} ${MLM_MODEL_CFG} | ||
| export MLM_EXTRA_ARGS="$*" |
Contributor
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
for script in \
tools/launcher/modules/Megatron-LM/examples/post_training/modelopt/train.sh \
tools/launcher/modules/Megatron-LM/examples/post_training/modelopt/finetune.sh
do
test -f "$script"
printf '\n== %s ==\n' "$script"
rg -n -C 8 'MLM_EXTRA_ARGS' "$script" || true
doneRepository: NVIDIA/Model-Optimizer
Length of output: 160
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== candidate files =='
git ls-files | rg '(^|/)(sft\.sh|train\.sh|finetune\.sh)$|MLM_EXTRA_ARGS' || true
printf '%s\n' '== sft.sh =='
sft='tools/launcher/common/megatron_lm/train/sft.sh'
test -f "$sft"
cat -n "$sft"
printf '%s\n' '== MLM_EXTRA_ARGS references =='
rg -n -C 10 'MLM_EXTRA_ARGS' . --glob '!node_modules/**' --glob '!dist/**' --glob '!build/**' || trueRepository: NVIDIA/Model-Optimizer
Length of output: 22438
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== repository references to Megatron-LM =='
rg -n -C 5 'Megatron-LM|post_training/modelopt/(train|finetune)\.sh|modules/' \
.gitmodules .gitignore README.md tools 2>/dev/null || true
printf '%s\n' '== possible submodule metadata =='
find . -maxdepth 4 \( -name .gitmodules -o -path '*/.git' \) -print
printf '%s\n' '== upstream raw scripts =='
base='https://raw.githubusercontent.com/NVIDIA/Model-Optimizer/main/tools/launcher/modules/Megatron-LM/examples/post_training/modelopt'
for name in train.sh finetune.sh; do
printf '\n-- %s --\n' "$name"
curl -L --fail --silent --show-error "$base/$name" | rg -n -C 12 'MLM_EXTRA_ARGS|argparse|python|train' || true
doneRepository: NVIDIA/Model-Optimizer
Length of output: 50381
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '%s\n' '== pinned Megatron-LM submodule =='
git config -f .gitmodules --get-regexp 'submodule\..*\.(path|url)' || true
git ls-tree HEAD tools/launcher/modules/Megatron-LM
printf '%s\n' '== exact upstream file probes =='
for ref in main $(git ls-tree HEAD tools/launcher/modules/Megatron-LM | awk '{print $3}'); do
[ -n "$ref" ] || continue
for name in train.sh finetune.sh; do
url="https://raw.githubusercontent.com/NVIDIA/Megatron-LM/$ref/examples/post_training/modelopt/$name"
body=$(curl -L --silent --show-error --write-out $'\nHTTP_STATUS:%{http_code}' "$url")
status=${body##*$'\nHTTP_STATUS:'}
content=${body%$'\nHTTP_STATUS:'*}
printf '\n-- ref=%s file=%s status=%s --\n' "$ref" "$name" "$status"
if [ "$status" = 200 ]; then
printf '%s\n' "$content" | rg -n -C 12 'MLM_EXTRA_ARGS|exec|python|ARGS|parse_args' || true
fi
done
done
printf '%s\n' '== launcher args handling =='
rg -n -C 8 'task\.args|args.*script|script.*args|sh .*script|subprocess|shell' \
tools/launcher --glob '*.py' --glob '*.md' | head -n 240Repository: NVIDIA/Model-Optimizer
Length of output: 27992
Preserve argument boundaries in MLM_EXTRA_ARGS.
The pinned train.sh and finetune.sh scripts expand ${MLM_EXTRA_ARGS} without quotes. Line 70 serializes positional parameters with "$*", so a value such as /tmp/data set.jsonl becomes separate arguments before the downstream script invokes Python. Use an argument-preserving interface or a documented serialization that the downstream scripts parse. Add a regression case for a whitespace-containing value.
🤖 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 `@tools/launcher/common/megatron_lm/train/sft.sh` at line 70, Update the
MLM_EXTRA_ARGS handling in the sft.sh argument flow to preserve positional
argument boundaries when values contain whitespace, rather than serializing them
with "$*". Use an argument-preserving interface or a serialization format that
the downstream train.sh and finetune.sh scripts explicitly parse, and add a
regression case covering a whitespace-containing value.
Comment on lines
+52
to
+54
| - MLM_MODEL_CFG: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | ||
| - QUANT_CFG: MAMBA_MOE_NVFP4_CONSERVATIVE_CFG | ||
| - MLM_MODEL_CKPT: /cicd/megatron-lm-bf16/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-MCore |
Contributor
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Use the requested aggressive NVFP4 configuration.
The PR objective requires MAMBA_MOE_NVFP4_AGGRESSIVE_CFG, but the workflow configures MAMBA_MOE_NVFP4_CONSERVATIVE_CFG. This produces a different PTQ artifact than the requested QAD workflow.
tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml#L52-L54: setQUANT_CFGtoMAMBA_MOE_NVFP4_AGGRESSIVE_CFG.tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml#L133-L135: setQUANT_CFGto the same aggressive configuration for export.
📍 Affects 1 file
tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml#L52-L54(this comment)tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml#L133-L135
🤖 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
`@tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml`
around lines 52 - 54, Update QUANT_CFG in
tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml
at lines 52-54 and 133-135 to use MAMBA_MOE_NVFP4_AGGRESSIVE_CFG consistently
for both QAD and export.
Comment on lines
+106
to
+118
| environment: | ||
| - MLM_MODEL_CFG: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | ||
| - MLM_MODEL_CKPT: /cicd/megatron-lm/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | ||
| - MLM_MODEL_SAVE: /cicd/megatron-lm-qad/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | ||
| - HF_MODEL_CKPT: /hf-local/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | ||
| - MLM_TRAIN_SCRIPT: finetune | ||
| - DATASET: nvidia/Nemotron-Post-Training-Dataset-v2 | ||
| - DP: "1" | ||
| - CP: "1" | ||
| - TP: "2" | ||
| - PP: "1" | ||
| - EP: "4" | ||
| - ETP: "1" |
Contributor
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Set QUANT_CFG for the QAD training wrapper.
task_2 invokes common/megatron_lm/train/sft.sh with --modelopt-enabled, but its environment does not set QUANT_CFG. Add the same NVFP4 configuration that the quantize and export stages use. Use MAMBA_MOE_NVFP4_AGGRESSIVE_CFG after the quantization configuration is corrected.
Proposed fix
environment:
- MLM_MODEL_CFG: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+ - QUANT_CFG: MAMBA_MOE_NVFP4_AGGRESSIVE_CFG
- MLM_MODEL_CKPT: /cicd/megatron-lm/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16Based on learnings, MLM_MODEL_CFG and QUANT_CFG are required for common/megatron_lm wrapper workflows. As per coding guidelines, “Set QUANT_CFG environment variable … when adding a new model config.”
📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| environment: | |
| - MLM_MODEL_CFG: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - MLM_MODEL_CKPT: /cicd/megatron-lm/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - MLM_MODEL_SAVE: /cicd/megatron-lm-qad/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - HF_MODEL_CKPT: /hf-local/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - MLM_TRAIN_SCRIPT: finetune | |
| - DATASET: nvidia/Nemotron-Post-Training-Dataset-v2 | |
| - DP: "1" | |
| - CP: "1" | |
| - TP: "2" | |
| - PP: "1" | |
| - EP: "4" | |
| - ETP: "1" | |
| environment: | |
| - MLM_MODEL_CFG: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - QUANT_CFG: MAMBA_MOE_NVFP4_AGGRESSIVE_CFG | |
| - MLM_MODEL_CKPT: /cicd/megatron-lm/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - MLM_MODEL_SAVE: /cicd/megatron-lm-qad/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - HF_MODEL_CKPT: /hf-local/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | |
| - MLM_TRAIN_SCRIPT: finetune | |
| - DATASET: nvidia/Nemotron-Post-Training-Dataset-v2 | |
| - DP: "1" | |
| - CP: "1" | |
| - TP: "2" | |
| - PP: "1" | |
| - EP: "4" | |
| - ETP: "1" |
🤖 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
`@tools/launcher/examples/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/megatron_lm_qad.yaml`
around lines 106 - 118, Update the environment block for the QAD training
workflow to define QUANT_CFG using the same corrected NVFP4 configuration as the
quantize and export stages, specifically MAMBA_MOE_NVFP4_AGGRESSIVE_CFG. Keep
the existing MLM_MODEL_CFG and other training settings unchanged.
Sources: Coding guidelines, Learnings
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #2134 +/- ##
==========================================
+ Coverage 78.73% 78.76% +0.03%
==========================================
Files 522 522
Lines 60342 60342
==========================================
+ Hits 47508 47527 +19
+ Misses 12834 12815 -19
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Contributor
|
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 does this PR do?
Type of change: New example
Add a Nemotron Nano 3 QAD Launcher Example on OSS Nemotron-Post-Training-V2 data. It performs 4 steps
MAMBA_MOE_NVFP4_AGGRESSIVE_CFGquant configchatdata. To train on a different subset or load the entire dataset, you may modify--finetune-data-splitand--finetune-data-filesflags.All steps use the TE (Transformer Engine) spec, which with the new TEGroupedMLP per-expert quantizer is approximately 10-15% faster than the previous local ModelOpt spec (which used SequentialMLP) on Hybrid-MoE models.
Usage
Testing
Before your PR is "Ready for review"
Make sure you read and follow Contributor guidelines and your commits are signed (
git commit -s -S).Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded
trust_remote_code=True,torch.load(..., weights_only=False),pickle, etc.).CONTRIBUTING.md: ✅ / ❌ / N/AAdditional Information
Summary by CodeRabbit
New Features
MLM_TRAIN_SCRIPT.Updates