update MI355 yaml for v26.5

[vidushi8]

No description provided.

[Copilot]

Pull request overview

Updates MI355/MI355X example training configurations for the v26.5 environment, primarily adjusting run hyperparameters and toggles to better match expected runtime behavior.

Changes:

• Adjusts batch sizes / train iteration counts across several MI355X Megatron and TorchTitan example configs.
• Tweaks profiling defaults in an MLPerf config and allocator behavior in an MLPerf launch script.
• Enables Transformer Engine cross-entropy fusion implementation in the MI355X Megatron GPT-OSS FP8 config.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
examples/torchtitan/configs/MI355X/deepseek_v3_16b-BF16-pretrain.yaml	Lowers TorchTitan local_batch_size for MI355X.
examples/mlperf/gpt_oss_20b/configs/MI355/gpt_oss_20B-FP8-mlperf-pretrain.yaml	Updates profiling env-var defaults (needs fix for boolean default casing).
examples/mlperf/gpt_oss_20b/config_MI355X_1x8x1_tp1pp1ep1_gbs32.sh	Disables PyTorch expandable segments allocator via env var.
examples/megatron/configs/MI355X/qwen3_30B_A3B-FP8-pretrain.yaml	Updates Megatron training iters and batch sizes for MI355X.
examples/megatron/configs/MI355X/qwen3_30B_A3B-BF16-pretrain.yaml	Updates Megatron training iters for MI355X.
examples/megatron/configs/MI355X/llama3.1_70B-FP8-pretrain.yaml	Adjusts Megatron micro/global batch sizes for MI355X.
examples/megatron/configs/MI355X/gpt_oss_20B-FP8-pretrain.yaml	Adjusts Megatron training iters/batches and selects TE fused cross-entropy implementation.
examples/megatron/configs/MI355X/gpt_oss_20B-BF16-pretrain.yaml	Updates Megatron training iters for MI355X.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

examples/mlperf/gpt_oss_20b/configs/MI355/gpt_oss_20B-FP8-mlperf-pretrain.yaml:51

• profile / use_pytorch_profiler are resolved via Primus env substitution after YAML parsing. The YAML loader only type-casts numbers (see primus/core/config/yaml_loader.py:_resolve_env_in_string), so ${PRIMUS_PROFILE:False} produces the string "False" (or an env-provided string), not a boolean. Megatron configs are then applied directly into the Megatron Namespace (primus/backends/megatron/megatron_adapter.py → MegatronArgBuilder.update), so a non-empty string here will be truthy and can unintentionally enable profiling.

To keep this env-toggle pattern safe with the current loader, use numeric 0/1 defaults (as this same file already documents for create_attention_mask_in_dataloader) or avoid env substitution for boolean fields.

      # profile
      profile: ${PRIMUS_PROFILE:False}
      use_pytorch_profiler: ${PRIMUS_PROFILE:False}