[hnyls2002]

• fused_mxfp8_rows in deepseek_v4_mhc.py decides when the mHC post and RMSNorm steps write MXFP8 activations directly for the next DeepGEMM consumer instead of writing bf16 and quantizing in a separate kernel.
• It previously returned False for every extend batch, so eager prefill always paid the separate quantize pass.
• Extend batches now take the fused path, except inside a breakable prefill graph, where the captured buffers stay bf16. Decode and verify keep the existing row cap, and sequence-parallel batches stay excluded.

No change to the computed values: the fused producer emits the same MXFP8 tensor the separate quantize kernel produced.

Verification: DeepSeek-V4.1-Flash DEP4 eager prefill is faster at large batches; the breakable-graph prefill path is unchanged; GSM8K is unchanged.

---

CI States

Latest PR Test (Base): 🚫 Run #39861312160
Latest PR Test (Extra): ❌ Run #39861312121
Latest PR Test (AMD ROCm 10): 🚫 Run #39861312192

[hnyls2002]

/rerun-test test_deepseek_v41_flash_gb300_dp2ep2.py test_deepseek_v41_flash_gb300_encoder_bcg.py

[github-actions]

🚀 2-gpu-gb300 (2 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300_dp2ep2.py
cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300_encoder_bcg.py

[hnyls2002]

/tag-and-rerun-ci