perf(hcu): optimize DeepSeek-V4 DSpark decode paths

[lungy666]

Motivation

DeepSeek-V4 Flash with DSpark on HCU could not use several optimized kernels
without broad global switches or a matching recent sgl-kernel wheel:

1. DsparkDraftSampler only admitted VanillaMarkov to the fused greedy
   proposal path. DSparkV4MarkovHead has the same vanilla Markov recurrence,
   but is a separate class and therefore materialized full-vocabulary bias and
   corrected-logit tensors on every proposal step.
2. Some released HCU images do not contain
   torch.ops.sgl_kernel.deepseek_v4_topk_transform_512, while their LightOp
   package exposes the ABI-compatible fused transform.
3. DSV4's existing SGLANG_OPT_BF16_FP32_GEMM_ALGO selector could not select
   the already-supported AITER path independently of the broad
   SGLANG_USE_AITER switch.

The result was avoidable decode latency on BW1000, especially at batch size 1.

Modifications

• Make DSpark fused-greedy dispatch capability-based instead of checking the
  concrete VanillaMarkov class.
• Implement sample_block_greedy_fused for DSparkV4MarkovHead using the
  existing MarkovGreedyStep kernel. It crops padded vocabulary columns and
  retains the eager path for genuinely TP-sharded Markov heads.
• Prefer the existing sgl-kernel DSV4 AOT top-k op on HIP, and fall back to
  LightOp's ABI-compatible fused top-k/page-table transform when that op is not
  present.
• Accept SGLANG_OPT_BF16_FP32_GEMM_ALGO=aiter as a narrow DSV4 compressor
  switch. SGLANG_USE_AITER=true keeps its existing behavior.
• Add a CPU unit test proving a non-VanillaMarkov head can opt into the fused
  greedy interface without entering the eager fallback.

All performance-sensitive choices remain opt-in except the missing-AOT-op
fallback. The default compressor algorithm remains cublas.

Accuracy Tests

Test system:

• 8 x Hygon BW1000 (gfx936), DTK 26.04, PyTorch 2.10
• DeepSeek-V4-Flash-0731-W4A8-INT4-Channel-Attn-W8A8-INT8-Channel
• TP8 + DP8 + DP Attention + DP LM head, DSpark gamma 5
• temperature 0

GSM8K, five-shot Chat API, 50 examples, 8 client threads, max 1024 output
tokens:

Score	Correct	Latency	Output throughput	Mean DSpark accept length
1.000	50/50	21.370 s	281.606 tok/s	4.1817

The fixed long-form Chinese prompt also produced coherent, non-repeating
technical prose for exactly 1024 output tokens.

Precision note: AITER's tuned BF16 GEMM returns BF16-rounded values before this
function converts the tensor back to FP32. For M=6, N=1024, K=4096, comparison
against torch.mm(..., out_dtype=torch.float32) measured max absolute error
0.498169 and mean absolute error 0.070678. For that reason the AITER algorithm
is explicitly opt-in; users needing the original FP32-accumulation path keep
the default cublas setting.

The fused Markov kernel can select a different draft token on rare near ties
because it accumulates in FP32 instead of reproducing eager BF16 intermediate
rounding. These are proposals only; target verification still controls emitted
tokens.

Speed Tests and Profiling

Single-request OpenAI Chat API benchmark with a fixed realistic technical
writing prompt, temperature=0, ignore_eos=true, and exactly 1024 generated
tokens:

Compressor path	Runs	End-to-end throughput
strict FP32 accumulation	2	40.083, 41.024 tok/s
AITER BF16 GEMM (opt-in)	3, including cold restart	74.089, 73.407, 75.452 tok/s

The W4A8 MoE, communication settings, prompt, and other service configuration
were held constant for this A/B comparison.

Kernel measurements on one BW1000:

Operation and shape	Before	After	Speedup
DSV4 top-k transform, M=6, L=4096, K=512	0.3575 ms	0.0111 ms	32.2x
compressor GEMM, M=6, N=1024, K=4096	0.6583 ms	0.0384 ms	17.1x
DSpark Markov greedy step, vocab=128800	0.7703 ms	0.1243 ms	6.2x

The top-k parity sweep covered mixed sequence lengths and confirmed that this
image had no DSV4 sgl-kernel AOT op, so the LightOp fallback was exercised
rather than merely imported.

Tests

• pytest -q test/registered/spec/dspark/test_dspark_draft_path_default.py::TestDsparkFusedGreedyRouting::test_custom_markov_head_can_provide_fused_greedy_sampler
  • 1 passed
• Direct HCU parity test for DSV4 fused greedy, including padded vocabulary
  cropping and TP-sharded fallback: passed.
• Direct HCU LightOp top-k parity over sequence lengths 127, 512, 1025 and
  4096: passed.
• Direct HCU SGLANG_OPT_BF16_FP32_GEMM_ALGO=aiter routing and output
  shape/dtype test: passed.
• Black 26.1.0, isort 7.0.0, Ruff 0.15.1, codespell 2.4.1, compileall, and
  git diff --check: passed.

Checklist

• Format code with the repository's pinned formatter versions.
• Add and run a focused unit test.
• Document the new algorithm value and numerical behavior inline.
• Provide accuracy and speed benchmark results.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.

[lungy666]

/tag-run-ci-label