[Elastic EP] Support Model Runner V2 (#53934)

Signed-off-by: almogtavor <almogtavor@gmail.com>
Signed-off-by: Almog Tavor <70065337+almogtavor@users.noreply.github.com>
This commit is contained in:
Almog Tavor
2026-09-29 00:20:40 +00:00
committed by GitHub
parent c68eb98b69
commit 171d1b9ef6
12 changed files with 171 additions and 80 deletions
+13 -9
View File
@@ -234,10 +234,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
# Persistent buffer for intermediate tensors (non-first PP ranks).
self.intermediate_tensors: IntermediateTensors | None = None
# Data parallelism.
self.dp_size = self.parallel_config.data_parallel_size
self.dp_rank = self.parallel_config.data_parallel_rank
# Dual batch overlap. Created in initialize_kv_cache(), once everything
# it runs the microbatched forward with exists.
self.ubatch_runner: UBatchRunner | None = None
@@ -509,9 +505,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
if self.is_pooling_model and self.is_last_pp_rank:
self.pooling_runner = PoolingRunner(self.model, self.vllm_config)
eplb_models_added |= self.eplb.maybe_register_model(
self.model, self.model_config, load_dummy_weights
self.model, self.model_config
)
self.eplb.maybe_start_async_loop(eplb_models_added)
self.eplb.maybe_start_async_loop(eplb_models_added, load_dummy_weights)
if not self.is_first_pp_rank:
# For non-first PP ranks, create intermediate tensors sized
@@ -526,6 +522,16 @@ class GPUModelRunner(LoRAModelRunnerMixin):
get_offloader().post_init()
@property
def dp_size(self) -> int:
# Not cached: elastic EP rewrites parallel_config in place on reconfigure.
return self.parallel_config.data_parallel_size
@property
def dp_rank(self) -> int:
# Not cached: elastic EP rewrites parallel_config in place on reconfigure.
return self.parallel_config.data_parallel_rank
def get_model(self) -> nn.Module:
return self.model
@@ -2337,9 +2343,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
def setup_eplb_from_mapping(
self, expanded_physical_to_logical: torch.Tensor
) -> None:
self.eplb.setup_from_mapping(
self.model, self.model_config, expanded_physical_to_logical
)
self.eplb.setup_from_mapping(self.model_config, expanded_physical_to_logical)
########### EPLB methods end ###########