[Core][PCP] Select MRV2 when PCP is enabled (#50034)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Jul 27, 2026 at 20:03 UTC
60417b4b744c371453eddcf5c8fa0f184418c957
1 file changed
+9
vllm/config/vllm.py
+9
@@ -557,6 +557,10 @@ class VllmConfig:
557
if use_v2_model_runner is not None:
558
return use_v2_model_runner
559
560
+ # PCP runtime support is implemented only by the V2 model runner.
561
+ if self.parallel_config.prefill_context_parallel_size > 1:
562
+ return True
563
+
564
# DSpark is implemented only by the V2 GPU model runner, and DeepSeek-V4
565
# is not otherwise a default-V2 architecture, so force V2 for it. If V2
566
# is unsupported for the rest of the config, _validate_v2_model_runner
@@ -1460,6 +1464,11 @@ class VllmConfig:
1464
1465
if self.use_v2_model_runner:
1466
self._validate_v2_model_runner()
1467
+ elif self.parallel_config.prefill_context_parallel_size > 1:
1468
+ raise ValueError(
1469
+ "Prefill context parallelism requires Model Runner V2. "
1470
+ "Remove VLLM_USE_V2_MODEL_RUNNER=0."
1471
+ )
1472
1473
# Re-compute compile ranges after platform-specific config updates
1474
# (e.g., XPU may lower max_num_batched_tokens when MLA is enabled)