Laguna xs dflash support (#41880)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.6
parent
3af561ec0a
commit
969fbfb4a9
@@ -1003,6 +1003,7 @@ class SpeculativeConfig:
|
||||
"kimi_k25",
|
||||
"minimax_m2",
|
||||
"gemma4",
|
||||
"laguna",
|
||||
]
|
||||
if (
|
||||
self.method in ("eagle3", "extract_hidden_states", "dflash")
|
||||
|
||||
@@ -39,7 +39,12 @@ from vllm.model_executor.model_loader.weight_utils import (
|
||||
default_weight_loader,
|
||||
maybe_remap_kv_scale_name,
|
||||
)
|
||||
from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
|
||||
from vllm.model_executor.models.interfaces import (
|
||||
EagleModelMixin,
|
||||
SupportsEagle3,
|
||||
SupportsLoRA,
|
||||
SupportsPP,
|
||||
)
|
||||
from vllm.model_executor.models.utils import (
|
||||
AutoWeightsLoader,
|
||||
PPMissingLayer,
|
||||
@@ -554,7 +559,7 @@ class LagunaDecoderLayer(nn.Module):
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class LagunaModel(nn.Module):
|
||||
class LagunaModel(nn.Module, EagleModelMixin):
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
|
||||
@@ -633,8 +638,17 @@ class LagunaModel(nn.Module):
|
||||
hidden_states = intermediate_tensors["hidden_states"]
|
||||
residual = intermediate_tensors["residual"]
|
||||
|
||||
for layer in islice(self.layers, self.start_layer, self.end_layer):
|
||||
aux_hidden_states = self._maybe_add_hidden_state(
|
||||
[], self.start_layer, hidden_states, residual
|
||||
)
|
||||
for layer_idx, layer in enumerate(
|
||||
islice(self.layers, self.start_layer, self.end_layer),
|
||||
start=self.start_layer,
|
||||
):
|
||||
hidden_states, residual = layer(positions, hidden_states, residual)
|
||||
self._maybe_add_hidden_state(
|
||||
aux_hidden_states, layer_idx + 1, hidden_states, residual
|
||||
)
|
||||
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors(
|
||||
@@ -642,6 +656,8 @@ class LagunaModel(nn.Module):
|
||||
)
|
||||
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
if len(aux_hidden_states) > 0:
|
||||
return hidden_states, aux_hidden_states
|
||||
return hidden_states
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
@@ -821,7 +837,7 @@ class LagunaModel(nn.Module):
|
||||
return loaded_params
|
||||
|
||||
|
||||
class LagunaForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
|
||||
class LagunaForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SupportsEagle3):
|
||||
fall_back_to_pt_during_load = False
|
||||
|
||||
packed_modules_mapping = {
|
||||
|
||||
Reference in New Issue
Block a user