cp: [recipe,model,ckpt] DeepSeek-V4 backports (4131, 4271, 4305, 4306, 4338) into r0.5.0

[cuichenx]

Summary

Combined backport to r0.5.0 for the requested DeepSeek-V4 changes.

Included cherry-picks:

• fix(model): keep full rotary_percent for DeepSeek-V4 MLA rope #4271 / 00174bc14952876c67b827a0b2c55457c3112617 -> d745e8b0d3eeb4af90eb21cb38745ca22b3a785a
• [recipe,docs,test] feat: add DeepSeek-V4-Flash SFT recipes, launcher and tests #4131 / 3af343edc83e0352ca8c54acfb94766340fbfa2d -> 8e100bd521c5e27c3bf0993dcb36dcb8ba626116
• feat(recipe): enable rope fusion in DeepSeek-V4-Flash SFT recipes #4305 / c9ab9cc124a055392927eee3b2220bc0049caa18 -> bc8692627c50e06632b499205153302cc4eeb3c0
• fix(deepseek): gate H100 fused kernel defaults #4338 / 5f2423312a7f19bebe0c881209900857b58b4a6c -> f7d51e4d782da155dba31c170e909aaea1719410
• fix(deepseek): gate H100 fused kernel defaults #4338 follow-up test fix / 655eb6d0b3b9a0c0ef31de234d2fdcabc7ad774a -> 99223662411ea036a9f72984f67ee843515e0ee7
• fix(ckpt): auto-generate DSv4 pipeline layout on export when run_config lacks it #4306 / ae87ec35e0dc14bb2c73a2156196a6e757c89f77 -> 8606ebc4e3c844f5cd3b88fa0a378cfc072b54b0

#4131 is included because #4305 updates the DeepSeek-V4-Flash SFT recipes introduced by #4131.

#4306 is included because the DSv4 PP>1 import/export round trip can otherwise fail when run_config.yaml serializes the finalized pipeline layout as an object stub instead of a concrete list. The backport regenerates and forwards the resolved layout during export.

The #4338 follow-up addresses the Claude bot comment about H100 CI determinism by forcing Blackwell fused-kernel support in the shared DeepSeek-V4 recipe test builder; the explicit unavailable-kernel test still overrides the helper to False.

Validation

• git diff --check origin/r0.5.0..HEAD
• uv run --no-sync python -m py_compile examples/conversion/convert_checkpoints_multi_gpu.py src/megatron/bridge/models/model_provider.py 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 tests/functional_tests/test_groups/recipes/test_deepseek_recipes_finetune.py
• uv run --no-sync --with pre-commit pre-commit run --all-files

Attempted targeted unit tests with uv run --frozen --group test python -m pytest tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py tests/unit_tests/recipes/test_deepseek_recipes.py, but this host cannot install the locked nvidia-resiliency-ext==0.6.0 wheel because the available wheel tag is manylinux_2_39_x86_64 while the host is manylinux_2_31_x86_64.

[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.