[Minor] Invert MRV1 specdec method fallback logic (#56899)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Sep 14, 2026 at 18:56 UTC
ceaa9f95ddbae0203ae444ec864bbbfb68b9b046
1 file changed
+10
-10
vllm/config/vllm.py
+10
-10
@@ -2829,16 +2829,16 @@ class VllmConfig:
2829
unsupported.append("pipeline parallelism with external_launcher")
2830
2831
if speculative_config is not None:
2832
- # TODO: ngram / ngram_gpu are not supported by the v2 model runner yet
2833
- if speculative_config.method in ("ngram", "ngram_gpu"):
2834
- unsupported.append("ngram/ngram_gpu speculative decoding")
2835
- elif speculative_config.method not in (
2836
- "eagle",
2837
- "eagle3",
2838
- "mtp",
2839
- "dflash",
2840
- "dspark",
2841
- "extract_hidden_states",
2832
+ if speculative_config.method in (
2833
+ # https://github.com/vllm-project/vllm/pull/40704
2834
+ "ngram",
2835
+ "ngram_gpu",
2836
+ # https://github.com/vllm-project/vllm/pull/43091
2837
+ "draft_model",
2838
+ "suffix",
2839
+ "medusa",
2840
+ "mlp_speculator",
2841
+ "custom_class",
2842
):
2843
unsupported.append(f"speculative method '{speculative_config.method}'")
2844