[Config] Remove Engram CUDA-alike device restrictions (#59171)
Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com> Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
This commit is contained in:
@@ -1588,14 +1588,14 @@ def test_engram_dp_shared_memory_config_validation(
|
||||
("DeepseekV41ForCausalLM", [1], "cuda", True),
|
||||
("DeepseekV41ForCausalLM", [1], "rocm", True),
|
||||
("DeepseekV41ForCausalLM", [], "cuda", False),
|
||||
("DeepseekV41ForCausalLM", [1], "cpu", False),
|
||||
("DeepseekV41ForCausalLM", [1], "cpu", True),
|
||||
("Qwen4ExpForCausalLM", [1], "cuda", True),
|
||||
("Qwen4ExpForCausalLM", [1], "rocm", True),
|
||||
("Qwen4ExpForConditionalGeneration", [1], "cuda", True),
|
||||
("Qwen4ExpForConditionalGeneration", [1], "rocm", True),
|
||||
("Qwen4ExpForCausalLM", [], "cuda", False),
|
||||
("Qwen4ExpForCausalLM", None, "cuda", False),
|
||||
("Qwen4ExpForCausalLM", [1], "cpu", False),
|
||||
("Qwen4ExpForCausalLM", [1], "cpu", True),
|
||||
("LlamaForCausalLM", [1], "cuda", False),
|
||||
("Qwen4ExpMTP", [], "cuda", False),
|
||||
(None, None, "cuda", False),
|
||||
|
||||
@@ -67,8 +67,6 @@ class EngramConfig:
|
||||
|
||||
def verify_model_config(self, model_config: "ModelConfig | None") -> None:
|
||||
"""Reject Engram configuration for models without n-gram embeddings."""
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
field = (
|
||||
_NGRAM_LAYER_FIELDS.get(model_config.architecture)
|
||||
if model_config is not None
|
||||
@@ -77,13 +75,11 @@ class EngramConfig:
|
||||
if (
|
||||
model_config is None
|
||||
or field is None
|
||||
or not current_platform.is_cuda_alike()
|
||||
or not getattr(model_config.hf_text_config, field, None)
|
||||
):
|
||||
raise ValueError(
|
||||
"EngramConfig requires a model with supported Engram "
|
||||
"embeddings, non-empty n-gram layer ids, and a CUDA-alike "
|
||||
"device (CUDA or ROCm)."
|
||||
"embeddings and non-empty n-gram layer ids."
|
||||
)
|
||||
|
||||
def resolve_dp_shared_memory(self, parallel_config: "ParallelConfig") -> None:
|
||||
|
||||
+1
-5
@@ -1370,8 +1370,6 @@ class VllmConfig:
|
||||
|
||||
def _resolve_and_verify_engram_config(self) -> None:
|
||||
"""Resolve defaults and validate n-gram embedding settings."""
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
model_config = self.model_config
|
||||
speculative_config = self.speculative_config
|
||||
# Draft configs inherit the target's communication groups and settings.
|
||||
@@ -1392,9 +1390,7 @@ class VllmConfig:
|
||||
"Disable --enable-dbo and set --ubatch-size to 0."
|
||||
)
|
||||
if self.engram_config is None:
|
||||
if not current_platform.is_cuda_alike() or not model_has_engram_layers(
|
||||
model_config
|
||||
):
|
||||
if not model_has_engram_layers(model_config):
|
||||
return
|
||||
self.engram_config = EngramConfig()
|
||||
self.engram_config.verify_model_config(model_config)
|
||||
|
||||
Reference in New Issue
Block a user