[YAMY1234]

Motivation

With SGLANG_SIMULATE_ACC_LEN set (simulated acceptance for benchmarking), DSpark already has three fast paths, but they are gated to HIP:

• the verify skips the draft / target accept sampling (softmax + rejection), whose result the simulated length overwrites;
• the accept and the sampled bonus fold into the verify CUDA graph (SGLANG_DSPARK_SIMULATE_FOLD);
• temperature-only rows are treated as greedy (SGLANG_SIMULATE_ACC_GREEDY).

On CUDA the same simulated runs still pay for the discarded sampling and the unfolded accept. This PR lets CUDA take the same paths.

Modifications

• dspark_verify.py: skip the accept sampling under simulated acceptance on every platform (was simulate and _is_hip).
• dspark_worker_v2.py: simulate_fold and _simulate_acc_greedy no longer require HIP; both keep their existing env switches (default on), so either can still be turned off.
• Remove the now-unused _is_hip / is_hip imports in both files.

No effect when SGLANG_SIMULATE_ACC_LEN is unset, so real serving is unchanged.

Accuracy Tests

Only the simulated-acceptance path changes. Accuracy runs use real acceptance (SGLANG_SIMULATE_ACC_LEN unset) and take the same code as before.

Speed Tests and Profiling

DeepSeek-V4.1-Flash, TP4 GB300, DSpark, --enable-nvidia-megakernel, AgentX agentic-coding, concurrency 1, --max-running-requests 1, SGLANG_SIMULATE_ACC_LEN=3.51 (match-expected, real-draft-token), 3,600 s. Measured on main 8e7601ae4a4 + #44945; the two arms differ only by this change (one boot each, different nodes):

	requests	tok/s/user p50	p90	TTFT p50
without this change	404	760.0	719.3	102 ms
with this change	474	949.7	926.0	96 ms

The two runs reached different context mixes (ISL p50 294K vs 153K), so here is the steady decode step from the server's decode-batch log, binned by context and matched on (conversation, turn):

context	≤ 64K	64–256K	> 256K
without this change	4.581 ms	4.595 ms	4.522 ms
with this change	3.654 ms	3.669 ms	3.723 ms

A rerun on current main will be added.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): 🚫 Run #39490479896
Latest PR Test (Extra): ❌ Run #39490483441
Latest PR Test (AMD ROCm 10): ⏳ Run #39490480327

[Oasis-Git]

[Codex]: Use /tag-and-rerun-ci for UT-only CI on current head f5a3cb8, as requested. Monitor the actual Stage A CPU and single-GPU small partitions and cancel this base workflow after Stage A passes. This does not validate all repository unit tests or full CI.

[Oasis-Git]

/tag-and-rerun-ci

[Oasis-Git]

[Codex]: UT-only rerun failed on f5a3cb8: Base attempt 3. Stage A CPU partitions 2 and 9 reproduced the same missing post_load_weights errors in the shared-expert-fusion and RunAI loader test doubles. The other 16 Stage A partitions retained their successful results. No further reruns or cancellation were performed; the workflow completed with failure.

[Oasis-Git]

[Codex]: UT Stage A passed on c5b9376: all 17 CPU partitions and the single GPU-small partition succeeded. Base run, attempt 1. The workflow was intentionally cancelled after Stage A, and completed cancellation is verified. Some Stage B jobs started before cancellation; this is not a full CI pass.