[Mypy] Fix typing for M models (#54262)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-09-01 11:26:00 -04:00
committed by GitHub
parent d9eb4e344f
commit 76f3249fbd
26 changed files with 512 additions and 200 deletions
+25 -4
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable
from typing import TYPE_CHECKING
import torch
import torch.nn as nn
@@ -33,6 +34,20 @@ from vllm.model_executor.models.utils import (
logger = init_logger(__name__)
if TYPE_CHECKING:
class _EagleMistralModelBase(nn.Module):
pass
class _EagleMistralForCausalLMBase(nn.Module):
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
raise NotImplementedError
else:
_EagleMistralModelBase = MistralModel
_EagleMistralForCausalLMBase = MistralForCausalLM
class EagleMistralDecoderLayer(MistralDecoderLayer):
def __init__(
self,
@@ -47,7 +62,7 @@ class EagleMistralDecoderLayer(MistralDecoderLayer):
@support_torch_compile
class EagleMistralModel(MistralModel):
class EagleMistralModel(_EagleMistralModelBase):
def __init__(
self,
*,
@@ -58,7 +73,9 @@ class EagleMistralModel(MistralModel):
# Bypass MistralModel.__init__ to avoid creating duplicate attention
# layer entries in the global context.
nn.Module.__init__(self)
self.config = vllm_config.speculative_config.draft_model_config.hf_config
speculative_config = vllm_config.speculative_config
assert speculative_config is not None
self.config = speculative_config.draft_model_config.hf_config
self.vocab_size = self.config.vocab_size
# Get drafter's quantization config
self.quant_config = get_draft_quant_config(vllm_config)
@@ -112,7 +129,9 @@ class EagleMistralModel(MistralModel):
return hidden_states, hidden_states
class EagleMistralForCausalLM(MistralForCausalLM, SupportsMultiModalEmbeddings):
class EagleMistralForCausalLM(
_EagleMistralForCausalLMBase, SupportsMultiModalEmbeddings
):
mistral_mapping = MistralForCausalLM.mistral_mapping | {
"eagle_linear": "model.fc",
}
@@ -121,7 +140,9 @@ class EagleMistralForCausalLM(MistralForCausalLM, SupportsMultiModalEmbeddings):
# Bypass MistralForCausalLM.__init__ to use the draft model config
# and to avoid creating an lm_head.
nn.Module.__init__(self)
self.config = vllm_config.speculative_config.draft_model_config.hf_config
speculative_config = vllm_config.speculative_config
assert speculative_config is not None
self.config = speculative_config.draft_model_config.hf_config
target_layer_num = vllm_config.model_config.get_num_layers(
vllm_config.parallel_config
)