A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 09:45 UTC
[Spec Decode] Don't fall back to V0 when spec decoding is enabled (#18265)
Woosuk Kwon committed
May 16, 2025 at 16:10 UTC
fabe89bbc41b34558558c0d5401e23776460eecb
1 file changed
+1
-9
vllm/engine/arg_utils.py
+1
-9
@@ -1325,7 +1325,7 @@ class EngineArgs:
1325
recommend_to_remove=False)
1326
return False
1327
1328
- # Only Ngram speculative decoding so far.
1328
+ # V1 supports N-gram, Medusa, and Eagle speculative decoding.
1329
is_ngram_enabled = False
1330
is_eagle_enabled = False
1331
is_medusa_enabled = False
@@ -1390,14 +1390,6 @@ class EngineArgs:
1390
_raise_or_fallback(feature_name=name, recommend_to_remove=False)
1391
return False
1392
1393
- # ngram is supported on V1, but off by default for now.
1394
- if is_ngram_enabled and _warn_or_fallback("ngram"):
1395
- return False
1396
-
1397
- # Eagle is under development, so we don't support it yet.
1398
- if is_eagle_enabled and _warn_or_fallback("Eagle"):
1399
- return False
1400
-
1393
# Non-[CUDA, TPU] may be supported on V1, but off by default for now.
1394
v0_hardware = not any(
1395
(current_platform.is_cuda(), current_platform.is_tpu()))