Hardware-agnostic model definition via HF transformer backend (1/N) (#49458)

Signed-off-by: Thomas Ortner <boh@zurich.ibm.com>
This commit is contained in:
Thomas Ortner
2026-08-13 11:21:50 +01:00
committed by GitHub
parent 903da602f3
commit b8baa31a28
10 changed files with 691 additions and 5 deletions
@@ -11,10 +11,7 @@ from torch import fx, nn
from transformers.activations import ACT2CLS
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import (
_ACTIVATION_AND_MUL_REGISTRY,
get_act_and_mul_fn,
)
from vllm.model_executor.layers.activation import _ACTIVATION_AND_MUL_REGISTRY
from vllm.model_executor.layers.linear import MergedColumnParallelLinear
from vllm.model_executor.models.transformers.fusers.base import StackedFuser
from vllm.model_executor.models.transformers.fx_utils import (
@@ -26,6 +23,7 @@ from vllm.model_executor.models.transformers.fx_utils import (
replace_expr,
single_self_call,
)
from vllm.model_executor.models.transformers.layers import get_act_and_mul_fn
from vllm.model_executor.models.transformers.utils import (
log_replacement,
replace_linear_class,
@@ -15,7 +15,6 @@ from vllm.distributed import (
)
from vllm.distributed.parallel_state import model_parallel_is_initialized
from vllm.distributed.utils import split_tensor_along_last_dim
from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm
from vllm.model_executor.models.transformers.fusers.base import BaseFuser
from vllm.model_executor.models.transformers.fx_utils import (
find_node,
@@ -25,6 +24,10 @@ from vllm.model_executor.models.transformers.fx_utils import (
peel,
trace,
)
from vllm.model_executor.models.transformers.layers import (
GemmaRMSNorm,
RMSNorm,
)
if TYPE_CHECKING:
from vllm.config import VllmConfig