[nvbug 5380101][fix] Fix nemotronNAS loading for TP>1 #6447

tomeras91 · 2025-07-29T10:59:16Z

Summary by CodeRabbit

New Features
- Improved support for distinguishing between key-value heads and attention heads in model configurations.
Bug Fixes
- Simplified internal logic for handling head dimensions and key-value head duplication in model weight mapping.
Tests
- Re-enabled a previously skipped integration test for automatic data type handling.

Description

This PR fixes weight loading for NemotronNAS with TP>1.

The issue was in layers where the number of KV heads is smaller than TP size. In this case, KV heads should be duplicated to match the TP size, so each TP rank can hold a full KV head. The logic to call _duplicate_kv where _num_kv_heads is a list was flawed.

The fix is to compute the number of KV heads from the shape of the K or V weight and the head dimension. The list logic that was present in _duplicate_kv_weights was added in a PR relevant only for Qwen3MOE, so it is now encapsulated only in Qwen3MoeHfWeightMapper.

Other than this, this PR also fixes the type hint for BaseWeightMapper._config.

Test Coverage

test_llm_api_pytorch.TestNemotronNas.test_auto_dtype_tp8 now passes

…d uses head_dim to get number of kv heads for the specific module Signed-off-by: Tomer Asida <[email protected]>

…en3 the special case. Now VGQA works for the general case and no need for special code for NemotronNAS Signed-off-by: Tomer Asida <[email protected]>

Signed-off-by: Tomer Asida <[email protected]>

…mpute num_kv_heads from head_dim and weight shape in standard flow Signed-off-by: Tomer Asida <[email protected]>

Signed-off-by: Tomer Asida <[email protected]>

coderabbitai · 2025-07-29T10:59:22Z

📝 Walkthrough

Walkthrough

The changes update type annotations and initialization logic in weight mapper classes, refine key-value head handling, and simplify key-value weight duplication logic. Additionally, a waiver for a specific integration test is removed, enabling its execution.

Changes

Cohort / File(s)	Change Summary
Base Weight Mapper Type Updates `tensorrt_llm/_torch/models/checkpoints/base_weight_mapper.py`	Refactored type annotations from `TConfig` to `ModelConfig` for attributes, method parameters, and property. Adjusted logic for setting `_head_dim` and related configuration handling in `init_model_and_config`.
Qwen3 MoE Weight Mapper Extension `tensorrt_llm/_torch/models/checkpoints/hf/qwen3_moe_weight_mapper.py`	Added/overrode `init_model_and_config` method to set `_num_kv_heads` based on model configuration, distinguishing between key-value and attention heads.
Key-Value Weight Duplication Logic `tensorrt_llm/_torch/models/checkpoints/hf/weight_mapper.py`	Simplified `_duplicate_kv_weights` by directly computing `num_kv_heads` from tensor shape and `_head_dim`, removing list-based logic and enumeration.
Test Waiver Removal `tests/integration/test_lists/waives.txt`	Removed skip entry for `accuracy/test_llm_api_pytorch.py::TestNemotronNas::test_auto_dtype_tp8`, re-enabling this test in integration runs.

Sequence Diagram(s)

sequenceDiagram
    participant Tester
    participant WaiverList

    Tester->>WaiverList: Check if test_auto_dtype_tp8 is skipped
    WaiverList-->>Tester: No longer skipped (entry removed)
    Tester->>Tester: Run test_auto_dtype_tp8 as part of integration suite

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~7 minutes

Suggested reviewers

yilin-void
brb-nv

Note

⚡️ Unit Test Generation is now available in beta!

Learn more here, or try it out under "Finishing Touches" below.

📜 Recent review details

Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between f35504f and 3ed48e0.

📒 Files selected for processing (1)

tests/integration/test_lists/waives.txt (0 hunks)

💤 Files with no reviewable changes (1)

tests/integration/test_lists/waives.txt

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

GitHub Check: Pre-commit Check

✨ Finishing Touches

🧪 Generate unit tests

Create PR with unit tests
Post copyable unit tests in a comment

🪧 Tips

Chat

There are 3 ways to chat with CodeRabbit:

Review comments: Directly reply to a review comment made by CodeRabbit. Example:
- I pushed a fix in commit <commit_id>, please review it.
- Explain this complex logic.
- Open a follow-up GitHub issue for this discussion.
Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
- @coderabbitai explain this code block.
- @coderabbitai modularize this function.
PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
- @coderabbitai gather interesting stats about this repository and render them as a table. Additionally, render a pie chart showing the language distribution in the codebase.
- @coderabbitai read src/utils.ts and explain its main purpose.
- @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.
- @coderabbitai help me debug CodeRabbit configuration file.

Support

Need help? Create a ticket on our support page for assistance with any issues or questions.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (Invoked using PR comments)

@coderabbitai pause to pause the reviews on a PR.
@coderabbitai resume to resume the paused reviews.
@coderabbitai review to trigger an incremental review. This is useful when automatic reviews are disabled for the repository.
@coderabbitai full review to do a full review from scratch and review all the files again.
@coderabbitai summary to regenerate the summary of the PR.
@coderabbitai generate docstrings to generate docstrings for this PR.
@coderabbitai generate sequence diagram to generate a sequence diagram of the changes in this PR.
@coderabbitai generate unit tests to generate unit tests for this PR.
@coderabbitai resolve resolve all the CodeRabbit review comments.
@coderabbitai configuration to show the current CodeRabbit configuration for the repository.
@coderabbitai help to get help.

Other keywords and placeholders

Add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.
Add @coderabbitai summary to generate the high-level summary at a specific location in the PR description.
Add @coderabbitai or @coderabbitai title anywhere in the PR title to generate the title automatically.

Documentation and Community

Visit our Documentation for detailed information on how to use CodeRabbit.
Join our Discord Community to get help, request features, and share feedback.
Follow us on X/Twitter for updates and announcements.

Copilot

Pull Request Overview

This PR fixes weight loading for NemotronNAS models when tensor parallelization (TP) is greater than 1, specifically addressing an issue where KV heads with fewer heads than TP size weren't being properly duplicated.

Fixes KV head duplication logic by computing heads from weight shapes instead of using a flawed list-based approach
Moves Qwen3MOE-specific logic into its dedicated weight mapper class
Corrects type hints for BaseWeightMapper configuration

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
tests/integration/test_lists/waives.txt	Removes test skip for NemotronNas TP8 test since it now passes
tensorrt_llm/_torch/models/checkpoints/hf/weight_mapper.py	Simplifies KV weight duplication by computing heads from weight shape
tensorrt_llm/_torch/models/checkpoints/hf/qwen3_moe_weight_mapper.py	Adds Qwen3MOE-specific KV heads initialization logic
tensorrt_llm/_torch/models/checkpoints/base_weight_mapper.py	Updates type hints and replaces num_kv_heads with head_dim calculation

tensorrt_llm/_torch/models/checkpoints/hf/qwen3_moe_weight_mapper.py

tensorrt_llm/_torch/models/checkpoints/base_weight_mapper.py

Signed-off-by: Tomer Asida <[email protected]>

tomeras91 · 2025-07-29T11:29:33Z

/bot run

tensorrt-cicd · 2025-07-29T11:34:34Z

PR_Github #13364 [ run ] triggered by Bot

shaharmor98

Short and sweet.
One little ask, approved anyway

tensorrt_llm/_torch/models/checkpoints/hf/weight_mapper.py

tensorrt-cicd · 2025-07-29T13:12:32Z

PR_Github #13364 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #9994 completed with status: 'FAILURE'

Signed-off-by: Tomer Asida <[email protected]>

tomeras91 · 2025-07-29T13:49:40Z

/bot run

tensorrt-cicd · 2025-07-29T13:55:07Z

PR_Github #13373 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-29T15:19:52Z

PR_Github #13373 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #10002 completed with status: 'FAILURE'

shaharmor98 · 2025-07-30T05:23:00Z

/bot run

tensorrt-cicd · 2025-07-30T05:28:16Z

PR_Github #13481 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-30T06:03:55Z

PR_Github #13481 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #10095 completed with status: 'FAILURE'

shaharmor98 · 2025-07-30T06:07:47Z

/bot run --disable-fail-fast

tensorrt-cicd · 2025-07-30T06:13:50Z

PR_Github #13491 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-30T11:22:18Z

PR_Github #13491 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #10103 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

Signed-off-by: Tomer Asida <[email protected]> Signed-off-by: Lanyu Liao <[email protected]>

Signed-off-by: Tomer Asida <[email protected]>

tomeras91 added 5 commits July 28, 2025 15:33

Add NemotronNASHFWeightMapper that overrides _duplicate_kv_weights an…

11a11b2

…d uses head_dim to get number of kv heads for the specific module Signed-off-by: Tomer Asida <[email protected]>

different fix: return to duplicating kv heads by head_dim and make Qw…

16f1011

…en3 the special case. Now VGQA works for the general case and no need for special code for NemotronNAS Signed-off-by: Tomer Asida <[email protected]>

fix typehint for BaseWeightMapper._config

9e8494e

Signed-off-by: Tomer Asida <[email protected]>

reduce code duplication - use the same _duplicate_kv function, but co…

e6c71c4

…mpute num_kv_heads from head_dim and weight shape in standard flow Signed-off-by: Tomer Asida <[email protected]>

remove failed test from waives.txt since it now passes

f6d0421

Signed-off-by: Tomer Asida <[email protected]>

tomeras91 requested a review from a team as a code owner July 29, 2025 10:59

tomeras91 requested review from dongxuy04 and nv-yilinf July 29, 2025 10:59

tomeras91 requested review from shaharmor98 and Copilot July 29, 2025 10:59

coderabbitai bot requested review from syuoni and yilin-void July 29, 2025 10:59

Copilot AI reviewed Jul 29, 2025

View reviewed changes

tensorrt_llm/_torch/models/checkpoints/hf/qwen3_moe_weight_mapper.py Show resolved Hide resolved

tensorrt_llm/_torch/models/checkpoints/base_weight_mapper.py Show resolved Hide resolved

Merge branch 'main' into fix-nemotron-nas-tp-gt-1

3e646eb

Signed-off-by: Tomer Asida <[email protected]>

coderabbitai bot requested a review from brb-nv July 29, 2025 11:28

shaharmor98 approved these changes Jul 29, 2025

View reviewed changes

tensorrt_llm/_torch/models/checkpoints/hf/weight_mapper.py Show resolved Hide resolved

Add explanation comment

f35504f

Signed-off-by: Tomer Asida <[email protected]>

coderabbitai bot requested a review from liji-nv July 29, 2025 14:10

shaharmor98 enabled auto-merge (squash) July 30, 2025 05:22

Merge branch 'main' into fix-nemotron-nas-tp-gt-1

3ed48e0

shaharmor98 merged commit a2514d9 into NVIDIA:main Jul 30, 2025
3 checks passed

tomeras91 deleted the fix-nemotron-nas-tp-gt-1 branch July 30, 2025 12:03

lancelly pushed a commit to lancelly/TensorRT-LLM that referenced this pull request Aug 6, 2025

[nvbug 5380101][fix] Fix nemotronNAS loading for TP>1 (NVIDIA#6447)

98d10df

Signed-off-by: Tomer Asida <[email protected]> Signed-off-by: Lanyu Liao <[email protected]>

jain-ria pushed a commit to jain-ria/TensorRT-LLM that referenced this pull request Aug 7, 2025

[nvbug 5380101][fix] Fix nemotronNAS loading for TP>1 (NVIDIA#6447)

42645e9

Signed-off-by: Tomer Asida <[email protected]>

[nvbug 5380101][fix] Fix nemotronNAS loading for TP>1 #6447

[nvbug 5380101][fix] Fix nemotronNAS loading for TP>1 #6447

Uh oh!

Conversation

tomeras91 commented Jul 29, 2025 • edited by coderabbitai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Description

Test Coverage

Uh oh!

coderabbitai bot commented Jul 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Suggested reviewers

Chat

Support

CodeRabbit Commands (Invoked using PR comments)

Other keywords and placeholders

Documentation and Community

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

Uh oh!

Uh oh!

Uh oh!

tomeras91 commented Jul 29, 2025

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

shaharmor98 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

tomeras91 commented Jul 29, 2025

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

shaharmor98 commented Jul 30, 2025

Uh oh!

tensorrt-cicd commented Jul 30, 2025

Uh oh!

tensorrt-cicd commented Jul 30, 2025

Uh oh!

shaharmor98 commented Jul 30, 2025

Uh oh!

tensorrt-cicd commented Jul 30, 2025

Uh oh!

tensorrt-cicd commented Jul 30, 2025

Uh oh!

Uh oh!

Uh oh!

tomeras91 commented Jul 29, 2025 •

edited by coderabbitai bot

Loading

coderabbitai bot commented Jul 29, 2025 •

edited

Loading