[https://nvbugs/6272666][fix] Added a GPU-count-aware post-processing block in pytorch_model_config.py that…

[tensorrt-cicd]

Summary

• Root cause: MiniMax-M2.5 FP8 8-GPU MoE ran as pure TP=8, sharding intermediate_size 1536→192 which violates the 128-granular FP8 block-scale requirement, aborting the shared MPI perf job; the titled "prefetch OOM" is a non-reproducing QA mislabel.
• Fix: Added a GPU-count-aware post-processing block in pytorch_model_config.py that, for minimax_m2.5_fp8 + gpus:8, sets moe_expert_parallel_size=8 and enable_attention_dp=True; removed the dead in-list ep:8 pattern block. Verified 2 passed, EXIT_CODE=0.
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6272666

Summary by CodeRabbit

Chores

• Updated Minimax-M2.5 FP8 model test configuration to optimize resource allocation and improve performance for multi-GPU distributed environments. Configuration refinements enhance consistency and efficiency in handling expert-based parallelization across GPU clusters, with improved support for distributed inference scenarios on large-scale GPU systems.

[coderabbitai]

Too many files changed? Review this PR in Change Stack to see how the pieces fit before you dive in.

📝 Walkthrough

Walkthrough

The PR relocates Minimax-M2.5 FP8 MoE configuration logic in the test model configuration file. Pattern-based routing that previously matched specific input/output dimensions is removed, and replaced with a post-pattern conditional override that applies MoE expert parallel sizing and attention data-parallel settings when the PyTorch backend runs the model on 8 GPUs.

Changes

Minimax-M2.5 FP8 Configuration Migration

Layer / File(s)	Summary
Minimax-M2.5 FP8 pattern-to-conditional configuration migration tests/integration/defs/perf/pytorch_model_config.py	Pattern-based pattern_configs entries for Minimax-M2.5 FP8 are removed. A new post-pattern conditional block is added that checks for PyTorch minimax_m2.5_fp8 on 8-GPU runs and applies moe_expert_parallel_size = 8 and enable_attention_dp = True to the base configuration.

Estimated Code Review Effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly Related PRs

• NVIDIA/TensorRT-LLM#14613: Prior work added pattern-based enable_attention_dp routing for the same minimax_m2.5_fp8 8-GPU configuration; this PR replaces that pattern-based approach with a post-pattern conditional override.

Suggested Reviewers

• leslie-fang25
• yufeiwu-nv
• tcherckez-nvidia

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is clear and specific, referencing the addition of a GPU-count-aware post-processing block in pytorch_model_config.py which matches the core change in the PR.
Description check	✅ Passed	The PR description includes root cause analysis, specific fix details, and test verification, but lacks structured sections matching the template (Description, Test Coverage, PR Checklist).
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

tests/integration/defs/perf/pytorch_model_config.py (1)

604-605: 💤 Low value

Consider using setdefault() for consistency.

The code directly assigns to base_config keys, which will unconditionally overwrite any values set by earlier pattern_configs. The DeepSeek FP8 block immediately above (lines 592-593) uses setdefault() to avoid clobbering existing config.

If this post-pattern override is intended to take precedence regardless of prior config, the current approach is correct. However, if you want to preserve values already set by pattern_configs, use setdefault().

♻️ Alternative using setdefault to preserve existing config

-    base_config['moe_expert_parallel_size'] = 8
-    base_config['enable_attention_dp'] = True
+    base_config.setdefault('moe_expert_parallel_size', 8)
+    base_config.setdefault('enable_attention_dp', True)

This would only set the values if they haven't already been configured by a pattern_configs entry.

🤖 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 `@tests/integration/defs/perf/pytorch_model_config.py` around lines 604 - 605,
Change the unconditional assignments to preserve prior pattern_configs by using
base_config.setdefault('moe_expert_parallel_size', 8) and
base_config.setdefault('enable_attention_dp', True) instead of directly
assigning to base_config['moe_expert_parallel_size'] and
base_config['enable_attention_dp']; this ensures existing values configured
earlier are not overwritten by these post-pattern defaults while keeping the
same defaults if keys are absent.

🤖 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 `@tests/integration/defs/perf/pytorch_model_config.py`:
- Line 603: The current substring check using model_label.lower() and 'gpus:8'
can match unintended labels like "gpus:80"; update the condition that decides
the 8-GPU config to perform an exact segment match instead of a raw
substring—e.g., normalize model_label, then either split the label on delimiters
(spaces, commas, slashes, colons) and compare tokens for the exact token
"gpus:8", or use a regex with anchors/boundaries to match
r'(?<!\d)gpus:8(?!\d)'; change the conditional that references model_label to
use this precise match so only true 8-GPU labels select the MiniMax-M2.5 FP8
configuration.

---

Nitpick comments:
In `@tests/integration/defs/perf/pytorch_model_config.py`:
- Around line 604-605: Change the unconditional assignments to preserve prior
pattern_configs by using base_config.setdefault('moe_expert_parallel_size', 8)
and base_config.setdefault('enable_attention_dp', True) instead of directly
assigning to base_config['moe_expert_parallel_size'] and
base_config['enable_attention_dp']; this ensures existing values configured
earlier are not overwritten by these post-pattern defaults while keeping the
same defaults if keys are absent.

🪄 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: 9b6b2a52-ef55-4b38-b93d-9c24016a8ded

📥 Commits

Reviewing files that changed from the base of the PR and between 86f9602 and 18834ee.

📒 Files selected for processing (1)

• tests/integration/defs/perf/pytorch_model_config.py

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Imprecise substring match for GPU count.

The substring check 'gpus:8' in model_label.lower() will match unintended labels like "gpus:80", "gpus:88", or any string containing "gpus:8" as a substring. This could incorrectly apply the 8-GPU MiniMax-M2.5 FP8 configuration to other GPU counts.

🔧 Proposed fix using more precise pattern matching

-    if ('pytorch' in model_label and 'minimax_m2.5_fp8' in model_label.lower()
-            and 'gpus:8' in model_label.lower()):
+    if ('pytorch' in model_label and 'minimax_m2.5_fp8' in model_label.lower()
+            and '-gpus:8-' in f'-{model_label.lower()}-'):
         base_config['moe_expert_parallel_size'] = 8
         base_config['enable_attention_dp'] = True

This wraps the label with delimiters to ensure "gpus:8" is an exact segment match, preventing false positives with "gpus:80" or "gpus:88".

🤖 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 `@tests/integration/defs/perf/pytorch_model_config.py` at line 603, The current
substring check using model_label.lower() and 'gpus:8' can match unintended
labels like "gpus:80"; update the condition that decides the 8-GPU config to
perform an exact segment match instead of a raw substring—e.g., normalize
model_label, then either split the label on delimiters (spaces, commas, slashes,
colons) and compare tokens for the exact token "gpus:8", or use a regex with
anchors/boundaries to match r'(?<!\d)gpus:8(?!\d)'; change the conditional that
references model_label to use this precise match so only true 8-GPU labels
select the MiniMax-M2.5 FP8 configuration.