[None][perf] Fuse KDA verify and opt-in prefill projections

[longlee0622]

Summary

• Reuse finalized fused KDA input-projection weights during speculative verify.
• Avoid six separate q/k/v/f/b/g GEMMs when fused weights are available.
• Add opt-in fused prefill projection via TLLM_KDA_ENABLE_FUSED_PREFILL_PROJ=1; default remains disabled for performance evaluation.
• Preserve the separate projection fallback when the env var is disabled or fused weights are unavailable.
• Cover fused prefill and fused verify against their sequential projection paths on B200.

Testing

• SKIP=ruff-format pre-commit run --files tensorrt_llm/_torch/models/modeling_kimi_linear.py tests/unittest/_torch/modeling/test_kimi_kda_fused_verify_parity.py
• python3 -m py_compile tensorrt_llm/_torch/models/modeling_kimi_linear.py tests/unittest/_torch/modeling/test_kimi_kda_fused_verify_parity.py
• B200 on umbriel-b200-027: python3 -m pytest --confcutdir=tests/unittest/_torch/modeling tests/unittest/_torch/modeling/test_kimi_kda_fused_verify_parity.py -v -x -s
  • Result: 2 passed
  • Fused prefill output, conv state, and SSM state: cosine 1.000000, relative L2 0
  • Verify round 1/2 output cosine: 0.999993
  • Verify committed-state cosine: 1.000000

The confcutdir flag isolates this target GPU test from the checkout-wide conftest. The available dev images prebuilt C++ bindings are from another branch and fail global collection before this Python-only test is reached.

Post-#17088 integration

The rebased prefill dispatch now preserves #17088 behavior in this order:

1. Full projection fusion when TLLM_KDA_ENABLE_FUSED_PREFILL_PROJ=1.
2. Otherwise use [TRTLLM-14703][fix] 1st cherrypick [combine MR99+MR100+MR102 on latest feat (with MR100 dispatch wiring)], 2nd cherrypick [deferred-finalize scales fix, fused-epilogue batch gate, hardening] #17088 qkvg_proj for default packed q/k/v when available.
3. Otherwise use separate q/k/v fallback.

Verify remains fully fused by default. Final head 7a35a5e022: local and GitHub pre-commit checks plus DCO pass. The B200 parity data above predates the rebase; the post-rebase change is dispatch integration and formatting, not kernel code.

[longlee0622]

Superseded by #17311