Stabilize Higgs Audio termination and prefill

[LiangSu8899]

Summary

• Kernelize the Higgs BF16 batched prefill path for embedding lookup, projection GEMMs, q/k norm + RoPE, and KV writes.
• Add a shared delayed-codebook helper for per-frame argmax, delay handling, and codebook embedding accumulation.
• Gate the delayed-codebook helper behind FLASHRT_ENABLE_AUDIO_CODEBOOK so source, binding, and required-symbol checks stay in lockstep.
• Fix delayed EOC handling so short BF16 prompts terminate instead of running to the generation cap.
• Make FP8 activation calibration independent of the first user request by using a fixed built-in calibration prompt.
• Add a deterministic repeated-code stop guard for pathological no-EOC FP8 loops.
• Refresh Higgs performance docs with the same single-stream benchmark methodology.

Root cause

The 40 s output cap was not a BF16 fallback issue. There were two related termination failures:

• BF16 batched prefill wrote a slightly different prompt KV state and the old delayed-EOC check only handled the first codebook reliably. Short prompts could miss EOC and run to the cap.
• FP8 used activation scales calibrated from the first user prompt. Different first prompts selected different scale sets; for some scale/prompt pairs the free-running discrete audio decode entered a stable repeated-code loop where EOC was no longer close to top-1.

The FP8 fix avoids first-request scale sensitivity and keeps a bounded deterministic termination guard for repeated code loops. Normal EOC termination remains unchanged.

Performance

Same single-stream Higgs benchmark setup as the docs.

Metric	Before	After
BF16 Four score. output duration	40.68 s cap	about 1 s
BF16 prefill, P=6	8.42 ms	6.73 ms
BF16 prefill, P=13	11.74 ms	6.82 ms
BF16 warm RTF, benchmark sentence	0.15	0.152
BF16 TTFA, benchmark sentence	~138 ms	127 ms
BF16 AR decode, benchmark sentence	~6.1 ms/frame	6.02 ms/frame
FP8 warm RTF, benchmark sentence	0.095-0.11	0.09
FP8 TTFA, benchmark sentence	~94 ms	79 ms
FP8 AR decode, benchmark sentence	~3.2 ms/frame	3.6 ms/frame

Known FP8 startup-order regressions were checked with first requests Four score., Four years ago., Hello from FlashRT., and the benchmark sentence; all subsequent Four score. / Four years ago. runs terminate normally.

The remaining torch matmul/embedding/argmax grep hits are limited to the one-time FP8 calibration reference path, not the BF16/FP8 request prefill/decode hot path.

Validation

• cmake -S . -B build -DGPU_ARCH=120 -DFLASHRT_ENABLE_AUDIO_CODEBOOK=ON
• cmake --build build --target flash_rt_kernels -j$(nproc)
• pytest -q tests/test_higgs_audio_v3_eoc.py tests/test_build_inventory.py tests/test_generic_kernel_bindings.py
• FLASHRT_BUILD_DIR=build-sm89-inventory pytest -q tests/test_build_inventory.py
• FLASHRT_BUILD_DIR=build-sm89-slim-inventory pytest -q tests/test_build_inventory.py
• FLASHRT_BUILD_DIR=build-sm89-audio-off-inventory pytest -q tests/test_build_inventory.py
• python -m py_compile flash_rt/frontends/torch/_higgs_audio_v3_fp8.py flash_rt/frontends/torch/_higgs_audio_v3_bf16.py flash_rt/frontends/torch/higgs_audio_v3_rtx.py
• git diff --check
• Import and required-symbol checks for Higgs kernel bindings
• Manual FP8 startup-order matrix for short-prompt termination
• Manual BF16 Higgs benchmark for Four score., Four years ago., and the documented benchmark sentence
• Diff-scoped scan for private paths, secrets, debug traces, and generated markers