fix(deepseek): gate H100 fused kernel defaults

[cuichenx]

Summary

• Add a DeepSeek V4 helper that enables Blackwell-only fused kernels only when CUDA capability is 10.x or newer.
• Use that helper for AutoBridge defaults so fused mHC and fused DSA default off on Hopper/H100 while staying on for Blackwell.
• Apply the same mHC gating in DeepSeek V4 recipes; DSA remains explicitly unfused in recipes and RoPE behavior is unchanged.

Validation

• H100 inference comparison on DFW, 2 nodes / 16 H100 GPUs, TP=1 PP=1 EP=16, Bridge bfb60bb8, MCore dev 2f1004963dcb1718804f3b858f5fb2fc73819694, NeMo 26.06 rc3 container:
  • 12760509, use_fused_mhc=False, apply_rope_fusion=False, apply_dsa_kernel_fusion=True: failed before model construction with AssertionError: apply_dsa_kernel_fusion requires SM100+ (Blackwell or later), but current device has compute capability 9.0.
  • 12760510, use_fused_mhc=False, apply_rope_fusion=False, apply_dsa_kernel_fusion=False: completed one-token inference, after_model: cuda_allocated_gib=47.57 cuda_reserved_gib=48.36, generated Hello2, exit 0:0.
• Containerized targeted tests passed: 58 tests from tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py and tests/unit_tests/recipes/test_deepseek_recipes.py in nvcr.io/nvidia/pytorch:26.04-py3.
• uv run --no-sync pre-commit run --all-files passed.
• uv run --no-sync ruff check src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py src/megatron/bridge/recipes/deepseek/deepseek_v4.py tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py tests/unit_tests/recipes/test_deepseek_recipes.py passed.
• uv run --no-sync ruff format --check src/megatron/bridge/models/deepseek/deepseek_v4_bridge.py src/megatron/bridge/recipes/deepseek/deepseek_v4.py tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py tests/unit_tests/recipes/test_deepseek_recipes.py passed.
• git diff --check passed.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[claude]

Code Review - [High] Existing recipe tests will fail on H100 CI runners. Location: tests/unit_tests/recipes/test_deepseek_recipes.py:315-318 (_build_deepseek_v4_recipe). Problem: _build_deepseek_v4_recipe monkeypatches AutoBridge but not deepseek_v4_supports_blackwell_fused_kernels. After this PR, use_fused_mhc is set dynamically by that helper. On H100 CI runners (sm_90, CUDA available), the helper returns False, so use_fused_mhc will be False -- but five existing tests assert it is True. Affected tests: test_deepseek_v4_adam_mxfp8_recipe_uses_validated_optimizer_defaults (line 342), test_deepseek_v4_muon_recipe_uses_validated_optimizer_defaults (line 372), test_deepseek_v4_base_recipe_uses_blackwell_defaults (line 387), test_deepseek_v4_flash_sft_recipe_uses_fused_mhc (line 423), test_deepseek_v4_flash_no_mtp_sft_recipe_disables_mtp (line 435). Suggested fix: Add monkeypatch.setattr for deepseek_v4_supports_blackwell_fused_kernels (lambda: True) inside _build_deepseek_v4_recipe. Suggested test cases: No perf tests impacted.

[Meirtz]

LGTM, thanks!