B300 NextN fp4: also honor speculative MoE runner backend in keep-decision (+ debug log)

The NextN/MTP draft is built by the EAGLE draft worker, where get_moe_runner_backend() didn't resolve
to flashinfer_trtllm, so the fp4 drop fired anyway -> 6144-vs-3072 crash in draft load. Check both the
main and speculative runner backends (spec is AUTO when --speculative-moe-runner-backend unset); keep
fp4 if either is a flashinfer fp4 backend. Temporary [NextN-fp4] warning logs the resolved backends.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/srt/models/deepseek_nextn.py

@@ -93,12 +93,28 @@ def _should_drop_nextn_modelopt_fp4_quant_config(
     """
     if quant_config is None or quant_config.get_name() != "modelopt_fp4":
         return False
-    backend = get_moe_runner_backend()
-    keep_fp4 = (
-        backend.is_flashinfer_trtllm()
-        or backend.is_flashinfer_trtllm_routed()
-        or backend.is_flashinfer_cutedsl()
-        or backend.is_flashinfer_cutlass()
+    # The NextN/MTP draft is built by the EAGLE draft worker, where the effective
+    # MoE runner may be the *speculative* backend (which is AUTO when
+    # --speculative-moe-runner-backend is unset), not the main runner. Keep fp4 if
+    # EITHER the main or the speculative runner is a flashinfer fp4 backend.
+    from sglang.srt.layers.moe.utils import get_speculative_moe_runner_backend
+
+    def _is_fp4_capable(b) -> bool:
+        return (
+            b.is_flashinfer_trtllm()
+            or b.is_flashinfer_trtllm_routed()
+            or b.is_flashinfer_cutedsl()
+            or b.is_flashinfer_cutlass()
+        )
+
+    main_backend = get_moe_runner_backend()
+    spec_backend = get_speculative_moe_runner_backend()
+    keep_fp4 = _is_fp4_capable(main_backend) or _is_fp4_capable(spec_backend)
+    logger.warning(
+        "[NextN-fp4] main_runner=%s spec_runner=%s keep_fp4=%s",
+        main_backend,
+        spec_backend,
+        keep_fp4,
     )
     return not keep_fp4