[alphabetc1]

Motivation

validate_hisparse accepts two combinations that crash the scheduler on the first request:

• Speculative decoding + HiSparse. The draft worker's HiSparseDSATokenToKVPool never gets a HiSparse mapping registered, so the first draft extend dies in translate_loc_to_hisparse_device:
  AttributeError: 'HiSparseDSATokenToKVPool' object has no attribute 'full_to_hisparse_device_index_mapping'.
• --enable-mixed-chunk + HiSparse. Mixing a prefill into the HiSparse decode batch fails in merge_batch:
  TypeError: can only concatenate list (not "NoneType") to list.

Both were reproduced on a 4-layer DeepSeek-V3.2 config with dummy weights (one H200, greedy, --disable-radix-cache). The control arms are identical to each other, so each crash is specific to the combination:

arm	result
no spec, no HiSparse	reference
HiSparse only	identical to reference
EAGLE/MTP only	identical to reference
EAGLE/MTP + HiSparse	scheduler crash on the first request
mixed chunk only (concurrent requests)	ok
mixed chunk + HiSparse	scheduler crash on the first mixed batch

Modifications

Reject both at startup with a clear error, until HiSparse supports them (speculative decoding is in progress in #40670 / #37771, which can drop the check).

Accuracy Tests

• With this change the speculative + HiSparse launch above exits at startup with ValueError: --enable-hisparse does not support speculative decoding yet.
• test_server_args.py, test_model_overrides.py: 385 passed.

Checklist

• Format your code according to the Format code with pre-commit.
• No new unit tests; this change only adds straightforward startup validation.

---

CI States

Latest PR Test (Base): ✅ Run #36293131226
Latest PR Test (Extra): ❌ Run #36293131028
Latest PR Test (AMD ROCm 10): ❌ Run #36293131140

[alphabetc1]

/tag-and-rerun-ci