Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -11,33 +11,41 @@ from transformers import GptOssConfig
|
||||
from vllm.attention import Attention, AttentionType
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import (get_ep_group, get_pp_group,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_gather)
|
||||
from vllm.distributed import (
|
||||
get_ep_group,
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_gather,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (QKVParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
ParallelLMHead, VocabParallelEmbedding)
|
||||
ParallelLMHead,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.model_executor.models.utils import sequence_parallel_chunk
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.utils import cdiv
|
||||
|
||||
from .interfaces import SupportsEagle3, SupportsPP
|
||||
from .utils import (AutoWeightsLoader, WeightsMapper, extract_layer_index,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory, make_layers,
|
||||
maybe_prefix)
|
||||
from .utils import (
|
||||
AutoWeightsLoader,
|
||||
WeightsMapper,
|
||||
extract_layer_index,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory,
|
||||
make_layers,
|
||||
maybe_prefix,
|
||||
)
|
||||
|
||||
|
||||
class OAIAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: GptOssConfig,
|
||||
@@ -59,16 +67,13 @@ class OAIAttention(nn.Module):
|
||||
base=config.rope_theta,
|
||||
dtype=torch.float32,
|
||||
rope_scaling={
|
||||
"rope_type":
|
||||
"yarn",
|
||||
"factor":
|
||||
config.rope_scaling["factor"],
|
||||
"original_max_position_embeddings":
|
||||
config.rope_scaling["original_max_position_embeddings"],
|
||||
"beta_fast":
|
||||
config.rope_scaling["beta_fast"],
|
||||
"beta_slow":
|
||||
config.rope_scaling["beta_slow"],
|
||||
"rope_type": "yarn",
|
||||
"factor": config.rope_scaling["factor"],
|
||||
"original_max_position_embeddings": config.rope_scaling[
|
||||
"original_max_position_embeddings"
|
||||
],
|
||||
"beta_fast": config.rope_scaling["beta_fast"],
|
||||
"beta_slow": config.rope_scaling["beta_slow"],
|
||||
},
|
||||
is_neox_style=True,
|
||||
)
|
||||
@@ -76,8 +81,8 @@ class OAIAttention(nn.Module):
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
|
||||
self.sinks = torch.nn.Parameter(
|
||||
torch.empty(config.num_attention_heads // tp_size,
|
||||
requires_grad=False))
|
||||
torch.empty(config.num_attention_heads // tp_size, requires_grad=False)
|
||||
)
|
||||
|
||||
self.q_size = self.num_attention_heads * self.head_dim // tp_size
|
||||
self.kv_size = self.num_key_value_heads * self.head_dim // tp_size
|
||||
@@ -104,8 +109,7 @@ class OAIAttention(nn.Module):
|
||||
self.num_local_key_value_heads = config.num_key_value_heads // tp_size
|
||||
|
||||
# Only apply sliding window to every other layer
|
||||
sliding_window = (config.sliding_window if self.layer_idx %
|
||||
2 == 0 else None)
|
||||
sliding_window = config.sliding_window if self.layer_idx % 2 == 0 else None
|
||||
self.attn = Attention(
|
||||
self.num_local_attention_heads,
|
||||
self.head_dim,
|
||||
@@ -119,8 +123,9 @@ class OAIAttention(nn.Module):
|
||||
sinks=self.sinks,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor,
|
||||
positions: torch.Tensor) -> torch.Tensor:
|
||||
def forward(
|
||||
self, hidden_states: torch.Tensor, positions: torch.Tensor
|
||||
) -> torch.Tensor:
|
||||
qkv, _ = self.qkv(hidden_states)
|
||||
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
||||
q, k = self.rotary_emb(positions, q, k)
|
||||
@@ -131,7 +136,6 @@ class OAIAttention(nn.Module):
|
||||
|
||||
|
||||
class MLPBlock(torch.nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
@@ -150,21 +154,22 @@ class MLPBlock(torch.nn.Module):
|
||||
self.num_experts = config.num_local_experts
|
||||
self.experts_per_token = config.num_experts_per_tok
|
||||
self.world_size = dist.get_world_size() if dist.is_initialized() else 1
|
||||
self.router = torch.nn.Linear(config.hidden_size,
|
||||
config.num_local_experts)
|
||||
self.router = torch.nn.Linear(config.hidden_size, config.num_local_experts)
|
||||
assert config.intermediate_size % self.world_size == 0
|
||||
self.experts = FusedMoE(num_experts=config.num_local_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
reduce_results=True,
|
||||
renormalize=True,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.experts",
|
||||
apply_router_weight_on_input=False,
|
||||
has_bias=True,
|
||||
activation="swigluoai",
|
||||
is_sequence_parallel=self.is_sequence_parallel)
|
||||
self.experts = FusedMoE(
|
||||
num_experts=config.num_local_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
reduce_results=True,
|
||||
renormalize=True,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.experts",
|
||||
apply_router_weight_on_input=False,
|
||||
has_bias=True,
|
||||
activation="swigluoai",
|
||||
is_sequence_parallel=self.is_sequence_parallel,
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
num_tokens = x.shape[0]
|
||||
@@ -181,7 +186,6 @@ class MLPBlock(torch.nn.Module):
|
||||
|
||||
|
||||
class TransformerBlock(torch.nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
@@ -193,12 +197,10 @@ class TransformerBlock(torch.nn.Module):
|
||||
cache_config = vllm_config.cache_config
|
||||
|
||||
self.layer_idx = extract_layer_index(prefix)
|
||||
self.attn = OAIAttention(config,
|
||||
prefix=f"{prefix}.attn",
|
||||
cache_config=cache_config)
|
||||
self.mlp = MLPBlock(vllm_config,
|
||||
self.layer_idx,
|
||||
prefix=f"{prefix}.mlp")
|
||||
self.attn = OAIAttention(
|
||||
config, prefix=f"{prefix}.attn", cache_config=cache_config
|
||||
)
|
||||
self.mlp = MLPBlock(vllm_config, self.layer_idx, prefix=f"{prefix}.mlp")
|
||||
self.input_layernorm = RMSNorm(config.hidden_size, eps=1e-5)
|
||||
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=1e-5)
|
||||
|
||||
@@ -213,19 +215,16 @@ class TransformerBlock(torch.nn.Module):
|
||||
residual = hidden_states
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
else:
|
||||
hidden_states, residual = self.input_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||
hidden_states = self.attn(hidden_states, positions)
|
||||
# Fully Connected
|
||||
hidden_states, residual = self.post_attention_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
|
||||
output = self.mlp(hidden_states)
|
||||
return output, residual
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class GptOssModel(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
@@ -249,9 +248,9 @@ class GptOssModel(nn.Module):
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
self.norm = RMSNorm(self.config.hidden_size, eps=1e-5)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], self.config.hidden_size))
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], self.config.hidden_size
|
||||
)
|
||||
self.aux_hidden_state_layers = tuple[int, ...]()
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
@@ -280,14 +279,10 @@ class GptOssModel(nn.Module):
|
||||
for i in range(self.start_layer, self.end_layer):
|
||||
layer = self.layers[i]
|
||||
if i in self.aux_hidden_state_layers:
|
||||
aux_hidden_states.append(x if residual is None else x +
|
||||
residual)
|
||||
aux_hidden_states.append(x if residual is None else x + residual)
|
||||
x, residual = layer(x, positions, residual)
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors({
|
||||
"hidden_states": x,
|
||||
"residual": residual
|
||||
})
|
||||
return IntermediateTensors({"hidden_states": x, "residual": residual})
|
||||
x, _ = self.norm(x, residual)
|
||||
|
||||
if len(aux_hidden_states) > 0:
|
||||
@@ -315,15 +310,12 @@ class GptOssModel(nn.Module):
|
||||
|
||||
intermediate_size = self.config.intermediate_size
|
||||
intermediate_size_block = intermediate_size // mxfp4_block
|
||||
per_rank_intermediate_size_block = cdiv(intermediate_size_block,
|
||||
tp_size)
|
||||
per_rank_intermediate_size = (per_rank_intermediate_size_block *
|
||||
mxfp4_block)
|
||||
per_rank_intermediate_size_block = cdiv(intermediate_size_block, tp_size)
|
||||
per_rank_intermediate_size = per_rank_intermediate_size_block * mxfp4_block
|
||||
|
||||
# Calculate common slicing bounds for current rank
|
||||
tp_rank_start = tp_rank * per_rank_intermediate_size
|
||||
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size,
|
||||
intermediate_size)
|
||||
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size, intermediate_size)
|
||||
|
||||
for name, weight in weights:
|
||||
# Skip layers on other devices.
|
||||
@@ -338,18 +330,17 @@ class GptOssModel(nn.Module):
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[:,
|
||||
2 * tp_rank_start:2 * tp_rank_end,
|
||||
...]
|
||||
narrow_weight = weight[:, 2 * tp_rank_start : 2 * tp_rank_end, ...]
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(
|
||||
param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None,
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif ".w2_weight_scale" in name:
|
||||
@@ -357,66 +348,68 @@ class GptOssModel(nn.Module):
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[..., tp_rank_start //
|
||||
mxfp4_block:tp_rank_end //
|
||||
mxfp4_block]
|
||||
narrow_weight = weight[
|
||||
..., tp_rank_start // mxfp4_block : tp_rank_end // mxfp4_block
|
||||
]
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(
|
||||
param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None,
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif ".w13_weight" in name:
|
||||
# Handle MLP gate and up projection weights
|
||||
# flat weight from (E, 2 * N, block_size, entry_per_block)
|
||||
# to (E, 2 * N, -1), shouldn't trigger copy for contiguous
|
||||
weight = weight.view(num_experts, 2 * intermediate_size,
|
||||
-1).contiguous()
|
||||
weight = weight.view(
|
||||
num_experts, 2 * intermediate_size, -1
|
||||
).contiguous()
|
||||
|
||||
# Extract gate and up projection parts
|
||||
# since the weight is shuffled, we can slice directly
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[:,
|
||||
2 * tp_rank_start:2 * tp_rank_end,
|
||||
...]
|
||||
narrow_weight = weight[:, 2 * tp_rank_start : 2 * tp_rank_end, ...]
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(
|
||||
param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None,
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif ".w2_weight" in name:
|
||||
# Handle MLP down projection weights
|
||||
# same flatten here, but since 2 mx4 value are packed in 1
|
||||
# uint8, divide by 2
|
||||
weight = weight.view(num_experts, -1,
|
||||
intermediate_size // 2).contiguous()
|
||||
weight = weight.view(
|
||||
num_experts, -1, intermediate_size // 2
|
||||
).contiguous()
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[...,
|
||||
tp_rank_start // 2:tp_rank_end // 2]
|
||||
narrow_weight = weight[..., tp_rank_start // 2 : tp_rank_end // 2]
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(
|
||||
param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None,
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif ".w13_bias" in name:
|
||||
@@ -425,35 +418,32 @@ class GptOssModel(nn.Module):
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[:,
|
||||
2 * tp_rank_start:2 * tp_rank_end]
|
||||
narrow_weight = weight[:, 2 * tp_rank_start : 2 * tp_rank_end]
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(
|
||||
param,
|
||||
narrow_weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None,
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif ".w2_bias" in name:
|
||||
# Handle MLP down projection bias
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
if use_ep:
|
||||
weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
# (only load on rank 0 to avoid duplication)
|
||||
if tp_rank != 0:
|
||||
weight.zero_()
|
||||
weight_loader(param,
|
||||
weight,
|
||||
weight_name=name,
|
||||
shard_id=None,
|
||||
expert_id=None)
|
||||
weight_loader(
|
||||
param, weight, weight_name=name, shard_id=None, expert_id=None
|
||||
)
|
||||
loaded_params.add(name)
|
||||
continue
|
||||
elif "sinks" in name:
|
||||
@@ -468,8 +458,7 @@ class GptOssModel(nn.Module):
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
if weight_loader == default_weight_loader:
|
||||
weight_loader(param, weight)
|
||||
else:
|
||||
@@ -480,8 +469,7 @@ class GptOssModel(nn.Module):
|
||||
if name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(param, weight)
|
||||
loaded_params.add(name)
|
||||
return loaded_params
|
||||
@@ -507,8 +495,7 @@ class GptOssModel(nn.Module):
|
||||
per_rank_intermediate_size = cdiv(intermediate_size, tp_size)
|
||||
# Calculate common slicing bounds for current rank
|
||||
tp_rank_start = tp_rank * per_rank_intermediate_size
|
||||
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size,
|
||||
intermediate_size)
|
||||
tp_rank_end = min((tp_rank + 1) * per_rank_intermediate_size, intermediate_size)
|
||||
|
||||
for name, weight in weights:
|
||||
# Skip layers on other devices.
|
||||
@@ -521,8 +508,7 @@ class GptOssModel(nn.Module):
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[:, :,
|
||||
2 * tp_rank_start:2 * tp_rank_end]
|
||||
narrow_weight = weight[:, :, 2 * tp_rank_start : 2 * tp_rank_end]
|
||||
|
||||
narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
|
||||
param = params_dict[name]
|
||||
@@ -548,8 +534,7 @@ class GptOssModel(nn.Module):
|
||||
if use_ep:
|
||||
narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
|
||||
else:
|
||||
narrow_weight = weight[:,
|
||||
2 * tp_rank_start:2 * tp_rank_end]
|
||||
narrow_weight = weight[:, 2 * tp_rank_start : 2 * tp_rank_end]
|
||||
|
||||
param = params_dict[name]
|
||||
param.copy_(narrow_weight)
|
||||
@@ -579,8 +564,7 @@ class GptOssModel(nn.Module):
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
if weight_loader == default_weight_loader:
|
||||
weight_loader(param, weight)
|
||||
else:
|
||||
@@ -591,14 +575,12 @@ class GptOssModel(nn.Module):
|
||||
if name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(param, weight)
|
||||
loaded_params.add(name)
|
||||
return loaded_params
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
(".qkv", ".q_proj", "q"),
|
||||
@@ -620,16 +602,29 @@ class GptOssModel(nn.Module):
|
||||
ep_rank_start = ep_rank * experts_per_rank
|
||||
ep_rank_end = (ep_rank + 1) * experts_per_rank
|
||||
|
||||
quant_method = (self.config.quantization_config['quant_method'] if
|
||||
hasattr(self.config, "quantization_config") else None)
|
||||
quant_method = (
|
||||
self.config.quantization_config["quant_method"]
|
||||
if hasattr(self.config, "quantization_config")
|
||||
else None
|
||||
)
|
||||
if quant_method == "mxfp4":
|
||||
return self._load_weights_mxfp4(ep_rank_end, ep_rank_start,
|
||||
heads_per_rank, head_start,
|
||||
weights, stacked_params_mapping)
|
||||
return self._load_weights_mxfp4(
|
||||
ep_rank_end,
|
||||
ep_rank_start,
|
||||
heads_per_rank,
|
||||
head_start,
|
||||
weights,
|
||||
stacked_params_mapping,
|
||||
)
|
||||
else:
|
||||
return self._load_weights_other(ep_rank_end, ep_rank_start,
|
||||
heads_per_rank, head_start,
|
||||
weights, stacked_params_mapping)
|
||||
return self._load_weights_other(
|
||||
ep_rank_end,
|
||||
ep_rank_start,
|
||||
heads_per_rank,
|
||||
head_start,
|
||||
weights,
|
||||
stacked_params_mapping,
|
||||
)
|
||||
|
||||
|
||||
class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
|
||||
@@ -641,17 +636,14 @@ class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
|
||||
},
|
||||
orig_to_new_suffix={
|
||||
".embed_tokens.weight": ".embedding.weight",
|
||||
|
||||
# MoE MXFP4 weights
|
||||
".gate_up_proj_blocks": ".w13_weight",
|
||||
".down_proj_blocks": ".w2_weight",
|
||||
".gate_up_proj_scales": ".w13_weight_scale",
|
||||
".down_proj_scales": ".w2_weight_scale",
|
||||
|
||||
# MoE other weights
|
||||
".gate_up_proj": ".w13_weight",
|
||||
".down_proj": ".w2_weight",
|
||||
|
||||
# MoE Bias
|
||||
".gate_up_proj_bias": ".w13_bias",
|
||||
".down_proj_bias": ".w2_bias",
|
||||
@@ -678,7 +670,8 @@ class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(self.config.vocab_size)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
self.model.make_empty_intermediate_tensors)
|
||||
self.model.make_empty_intermediate_tensors
|
||||
)
|
||||
|
||||
def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
|
||||
self.model.aux_hidden_state_layers = layers
|
||||
@@ -690,23 +683,22 @@ class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.get_input_embeddings(input_ids)
|
||||
|
||||
def forward(self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
return self.model(input_ids, positions, intermediate_tensors,
|
||||
inputs_embeds)
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
return self.model(input_ids, positions, intermediate_tensors, inputs_embeds)
|
||||
|
||||
def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
logits = self.logits_processor(self.lm_head, hidden_states)
|
||||
return logits
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
loader = AutoWeightsLoader(
|
||||
self,
|
||||
skip_prefixes=(["lm_head."]
|
||||
if self.config.tie_word_embeddings else None),
|
||||
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
|
||||
)
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
Reference in New Issue
Block a user