wehub-resource-sync
535 lines
19 KiB
Python
535 lines
19 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
from collections.abc import Iterable
|
|
from itertools import islice
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers import CohereConfig
|
|
|
|
from vllm.compilation.decorators import support_torch_compile
|
|
from vllm.config import CacheConfig, VllmConfig
|
|
from vllm.distributed import (
|
|
get_pp_group,
|
|
get_tensor_model_parallel_world_size,
|
|
)
|
|
from vllm.model_executor.layers.activation import SiluAndMul
|
|
from vllm.model_executor.layers.attention import Attention
|
|
from vllm.model_executor.layers.fused_moe import (
|
|
FusedMoE,
|
|
)
|
|
from vllm.model_executor.layers.linear import (
|
|
MergedColumnParallelLinear,
|
|
QKVParallelLinear,
|
|
ReplicatedLinear,
|
|
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 VocabParallelEmbedding
|
|
from vllm.model_executor.model_loader.weight_utils import (
|
|
row_parallel_weight_loader,
|
|
)
|
|
from vllm.model_executor.utils import set_weight_attrs
|
|
from vllm.platforms import current_platform
|
|
from vllm.sequence import IntermediateTensors
|
|
|
|
from .commandr import LayerNorm
|
|
from .interfaces import SupportsPP, SupportsQuant
|
|
from .utils import (
|
|
AutoWeightsLoader,
|
|
WeightsMapper,
|
|
extract_layer_index,
|
|
make_empty_intermediate_tensors_factory,
|
|
make_layers,
|
|
maybe_prefix,
|
|
)
|
|
|
|
|
|
def is_prefix_dense_layer(config: CohereConfig, layer_idx: int) -> bool:
|
|
"""True when layer_idx lies in the contiguous dense MLP prefix."""
|
|
if layer_idx >= len(config.mlp_layer_types):
|
|
return False
|
|
return all(t == "dense" for t in config.mlp_layer_types[: layer_idx + 1])
|
|
|
|
|
|
@torch.compile(backend=current_platform.simple_compile_backend)
|
|
def token_choice_with_bias(
|
|
hidden_states: torch.Tensor,
|
|
gating_output: torch.Tensor,
|
|
topk: int,
|
|
renormalize: bool,
|
|
):
|
|
"""Sigmoid -> top-k (-> renormalize) custom routing for Cohere2Moe."""
|
|
assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
|
|
|
|
scores = gating_output.float().sigmoid()
|
|
topk_weights, topk_ids = torch.topk(scores, k=topk, dim=-1, sorted=False)
|
|
|
|
if renormalize:
|
|
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
|
|
|
return topk_weights.to(torch.float32), topk_ids.to(torch.int32)
|
|
|
|
|
|
@torch.compile(backend=current_platform.simple_compile_backend)
|
|
def rms_norm_func(hidden_states, weight, variance_epsilon):
|
|
input_dtype = hidden_states.dtype
|
|
hidden_states = hidden_states.to(torch.float32)
|
|
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
|
hidden_states = hidden_states * torch.rsqrt(variance + variance_epsilon)
|
|
hidden_states = weight.to(torch.float32) * hidden_states
|
|
return hidden_states.to(input_dtype)
|
|
|
|
|
|
class RMSNorm(nn.Module):
|
|
def __init__(self, param_shape=None, eps=1e-6):
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(param_shape))
|
|
self.variance_epsilon = eps
|
|
set_weight_attrs(self.weight, {"weight_loader": row_parallel_weight_loader})
|
|
|
|
def forward(self, hidden_states, residuals=None):
|
|
hidden_states = rms_norm_func(hidden_states, self.weight, self.variance_epsilon)
|
|
return hidden_states, residuals
|
|
|
|
|
|
def select_norm_impl(config: CohereConfig) -> tuple[type[nn.Module], float]:
|
|
"""Returns (norm_class, eps). Uses RMSNorm when config.rms_norm_eps is set,
|
|
otherwise falls back to LayerNorm with config.layer_norm_eps."""
|
|
rms_eps = getattr(config, "rms_norm_eps", None)
|
|
if rms_eps is not None:
|
|
return RMSNorm, rms_eps
|
|
return LayerNorm, config.layer_norm_eps
|
|
|
|
|
|
class Cohere2MoeMLP(nn.Module):
|
|
"""Cohere MLP used as shared experts in the MoE block."""
|
|
|
|
def __init__(
|
|
self,
|
|
config: CohereConfig,
|
|
intermediate_size: int | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
reduce_results: bool = False,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.config = config
|
|
self.hidden_size = config.hidden_size
|
|
self.intermediate_size = (
|
|
intermediate_size
|
|
if intermediate_size is not None
|
|
else config.intermediate_size
|
|
)
|
|
self.gate_up_proj = MergedColumnParallelLinear(
|
|
self.hidden_size,
|
|
[self.intermediate_size] * 2,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.gate_up_proj",
|
|
)
|
|
self.down_proj = RowParallelLinear(
|
|
self.intermediate_size,
|
|
self.hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
reduce_results=reduce_results,
|
|
prefix=f"{prefix}.down_proj",
|
|
)
|
|
self.act_fn = SiluAndMul()
|
|
|
|
def forward(self, x):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.down_proj(x)
|
|
return x
|
|
|
|
|
|
class Cohere2MoeAttention(nn.Module):
|
|
"""Cohere MoE attention with sliding-window interleave."""
|
|
|
|
def __init__(
|
|
self,
|
|
config: CohereConfig,
|
|
cache_config: CacheConfig | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
self.config = config
|
|
self.layer_idx = extract_layer_index(prefix)
|
|
self.hidden_size = config.hidden_size
|
|
self.total_num_heads = config.num_attention_heads
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.head_dim = getattr(
|
|
config, "head_dim", self.hidden_size // self.total_num_heads
|
|
)
|
|
self.total_num_kv_heads = config.num_key_value_heads
|
|
if self.total_num_kv_heads >= tp_size:
|
|
assert self.total_num_kv_heads % tp_size == 0
|
|
else:
|
|
assert tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
|
self.q_size = self.num_heads * self.head_dim
|
|
self.kv_size = self.num_kv_heads * self.head_dim
|
|
self.scaling = self.head_dim**-0.5
|
|
self.max_position_embeddings = getattr(
|
|
config, "model_max_length", None
|
|
) or getattr(config, "max_position_embeddings", 8192)
|
|
self.qkv_proj = QKVParallelLinear(
|
|
self.hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.qkv_proj",
|
|
)
|
|
self.o_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
self.hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.o_proj",
|
|
)
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
max_position=self.max_position_embeddings,
|
|
rope_parameters=config.rope_parameters,
|
|
is_neox_style=False,
|
|
)
|
|
|
|
self.sliding_window = None
|
|
layer_types = getattr(config, "layer_types", None)
|
|
if (
|
|
layer_types is not None
|
|
and layer_types[self.layer_idx] == "sliding_attention"
|
|
):
|
|
self.sliding_window = config.sliding_window
|
|
|
|
# Prefix-dense layers have full attention (no sliding window). When
|
|
# prefix_dense_sliding_window_pattern == 1, they keep RoPE even though
|
|
# they are not sliding-window layers.
|
|
prefix_dense_sliding_window_pattern = getattr(
|
|
config, "prefix_dense_sliding_window_pattern", 1
|
|
)
|
|
self.force_rope = bool(
|
|
is_prefix_dense_layer(config, self.layer_idx)
|
|
and prefix_dense_sliding_window_pattern == 1
|
|
)
|
|
|
|
self.attn = Attention(
|
|
self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
per_layer_sliding_window=self.sliding_window,
|
|
prefix=f"{prefix}.attn",
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.qkv_proj(hidden_states)
|
|
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
|
if self.sliding_window or self.force_rope:
|
|
q, k = self.rotary_emb(positions, q, k)
|
|
attn_output = self.attn(q, k, v)
|
|
output, _ = self.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class Cohere2Moe(nn.Module):
|
|
"""Tensor-parallel MoE block for Cohere2Moe with shared experts."""
|
|
|
|
def __init__(
|
|
self,
|
|
config: CohereConfig,
|
|
params_dtype: torch.dtype | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
tp_size: int | None = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
self.tp_size = get_tensor_model_parallel_world_size()
|
|
|
|
if self.tp_size > config.num_experts:
|
|
raise ValueError(
|
|
f"Tensor parallel size {self.tp_size} is greater than "
|
|
f"the number of experts {config.num_experts}."
|
|
)
|
|
|
|
if (
|
|
hasattr(config, "expert_selection_fn")
|
|
and config.expert_selection_fn == "sigmoid"
|
|
):
|
|
self.custom_routing_function = token_choice_with_bias
|
|
else:
|
|
self.custom_routing_function = None
|
|
|
|
self.gate = ReplicatedLinear(
|
|
config.hidden_size,
|
|
config.num_experts,
|
|
bias=False,
|
|
params_dtype=params_dtype,
|
|
quant_config=None,
|
|
prefix=f"{prefix}.gate",
|
|
)
|
|
|
|
if hasattr(config, "num_shared_experts") and config.num_shared_experts > 0:
|
|
self.shared_experts = Cohere2MoeMLP(
|
|
config=config,
|
|
intermediate_size=config.intermediate_size * config.num_shared_experts,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.shared_experts",
|
|
)
|
|
self.shared_expert_combination_strategy = getattr(
|
|
config, "shared_expert_combination_strategy", "sum"
|
|
)
|
|
assert self.shared_expert_combination_strategy in ("average", "sum"), (
|
|
"shared_expert_combination_strategy must be one of ['average', 'sum']"
|
|
)
|
|
else:
|
|
self.shared_experts = None
|
|
self.shared_expert_combination_strategy = None
|
|
|
|
self.experts = FusedMoE(
|
|
num_experts=config.num_experts,
|
|
top_k=config.num_experts_per_tok,
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
params_dtype=params_dtype,
|
|
renormalize=getattr(config, "norm_topk_prob", True),
|
|
quant_config=quant_config,
|
|
tp_size=tp_size,
|
|
prefix=f"{prefix}.experts",
|
|
custom_routing_function=self.custom_routing_function,
|
|
shared_experts=self.shared_experts,
|
|
)
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
orig_shape = hidden_states.shape
|
|
hidden_states = hidden_states.view(-1, self.hidden_size)
|
|
router_logits, _ = self.gate(hidden_states)
|
|
# FusedMoE handles shared expert overlap internally and returns
|
|
# shared_output + routed_output when shared_experts is set.
|
|
final_hidden_states = self.experts(hidden_states, router_logits)
|
|
if self.shared_expert_combination_strategy == "average":
|
|
final_hidden_states = final_hidden_states / 2
|
|
return final_hidden_states.view(orig_shape)
|
|
|
|
|
|
class Cohere2MoeDecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: CohereConfig,
|
|
cache_config: CacheConfig | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.config = config
|
|
self.hidden_size = config.hidden_size
|
|
self.layer_idx = extract_layer_index(prefix)
|
|
|
|
self.self_attn = Cohere2MoeAttention(
|
|
config,
|
|
cache_config,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.self_attn",
|
|
)
|
|
|
|
if config.mlp_layer_types[self.layer_idx] == "dense":
|
|
self.mlp = Cohere2MoeMLP(
|
|
config=config,
|
|
intermediate_size=getattr(
|
|
config, "prefix_dense_intermediate_size", config.intermediate_size
|
|
),
|
|
quant_config=quant_config,
|
|
reduce_results=True,
|
|
prefix=f"{prefix}.mlp",
|
|
)
|
|
else:
|
|
self.mlp = Cohere2Moe(
|
|
config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
|
|
)
|
|
|
|
norm_cls, norm_eps = select_norm_impl(config)
|
|
self.input_layernorm = norm_cls(param_shape=(config.hidden_size,), eps=norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
residual: torch.Tensor | None,
|
|
) -> tuple[torch.Tensor, torch.Tensor]:
|
|
residual = hidden_states
|
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
|
hidden_states_attention = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
)
|
|
hidden_states_mlp = self.mlp(hidden_states)
|
|
|
|
hidden_states = residual + hidden_states_attention + hidden_states_mlp
|
|
return hidden_states, residual
|
|
|
|
|
|
@support_torch_compile
|
|
class Cohere2MoeModel(nn.Module):
|
|
"""Transformer decoder for Cohere2Moe."""
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
super().__init__()
|
|
|
|
config = vllm_config.model_config.hf_config
|
|
cache_config = vllm_config.cache_config
|
|
quant_config = vllm_config.quant_config
|
|
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.vocab_size = config.vocab_size
|
|
self.org_vocab_size = config.vocab_size
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size, config.hidden_size
|
|
)
|
|
|
|
# Decoder layers read per-layer MLP layout from config.mlp_layer_types
|
|
# (dense MLP vs MoE) and use it for weight loading. Transformers >=5.10
|
|
# populates this field; older versions only expose first_k_dense_replace.
|
|
# Normalize here so layer construction below sees a consistent layout.
|
|
if getattr(config, "mlp_layer_types", None) is None:
|
|
first_k_dense_replace = getattr(config, "first_k_dense_replace", None)
|
|
n = config.num_hidden_layers
|
|
if first_k_dense_replace is not None:
|
|
config.mlp_layer_types = ["dense"] * first_k_dense_replace + [
|
|
"sparse"
|
|
] * (n - first_k_dense_replace)
|
|
else:
|
|
config.mlp_layer_types = ["sparse"] * n
|
|
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda prefix: Cohere2MoeDecoderLayer(
|
|
config, cache_config, quant_config, prefix=prefix
|
|
),
|
|
prefix=f"{prefix}.layers",
|
|
)
|
|
norm_cls, norm_eps = select_norm_impl(config)
|
|
self.norm = norm_cls(param_shape=(config.hidden_size,), eps=norm_eps)
|
|
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
|
["hidden_states", "residual"], config.hidden_size
|
|
)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.embed_tokens(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None = None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
) -> torch.Tensor | IntermediateTensors:
|
|
if get_pp_group().is_first_rank:
|
|
if inputs_embeds is not None:
|
|
hidden_states = inputs_embeds
|
|
else:
|
|
hidden_states = self.get_input_embeddings(input_ids)
|
|
residual = None
|
|
else:
|
|
assert intermediate_tensors is not None
|
|
hidden_states = intermediate_tensors["hidden_states"]
|
|
residual = intermediate_tensors["residual"]
|
|
for layer in islice(self.layers, self.start_layer, self.end_layer):
|
|
hidden_states, residual = layer(positions, hidden_states, residual)
|
|
if not get_pp_group().is_last_rank:
|
|
return IntermediateTensors(
|
|
{"hidden_states": hidden_states, "residual": residual}
|
|
)
|
|
hidden_states, _ = self.norm(hidden_states, residual)
|
|
return hidden_states
|
|
|
|
|
|
class Cohere2MoeForCausalLM(nn.Module, SupportsPP, SupportsQuant):
|
|
is_text_generation_model = True
|
|
|
|
hf_to_vllm_mapper = WeightsMapper(
|
|
orig_to_new_stacked={
|
|
# weight_name: (param_name, shard_id)
|
|
".q_proj": (".qkv_proj", "q"),
|
|
".k_proj": (".qkv_proj", "k"),
|
|
".v_proj": (".qkv_proj", "v"),
|
|
# .experts.gate_up_proj must be handled by MoERunner.load_weights for EP
|
|
".mlp.gate_proj": (".mlp.gate_up_proj", 0),
|
|
".mlp.up_proj": (".mlp.gate_up_proj", 1),
|
|
".shared_experts.gate_proj": (".shared_experts.gate_up_proj", 0),
|
|
".shared_experts.up_proj": (".shared_experts.gate_up_proj", 1),
|
|
}
|
|
)
|
|
packed_modules_mapping = {
|
|
"qkv_proj": [
|
|
"q_proj",
|
|
"k_proj",
|
|
"v_proj",
|
|
],
|
|
"gate_up_proj": [
|
|
"gate_proj",
|
|
"up_proj",
|
|
],
|
|
}
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
super().__init__()
|
|
config = vllm_config.model_config.hf_config
|
|
quant_config = vllm_config.quant_config
|
|
self.config = config
|
|
assert getattr(config, "tie_word_embeddings", True)
|
|
self.unpadded_vocab_size = config.vocab_size
|
|
self.quant_config = quant_config
|
|
self.logits_scale = config.logit_scale
|
|
self.logits_processor = LogitsProcessor(
|
|
self.unpadded_vocab_size, config.vocab_size, scale=self.logits_scale
|
|
)
|
|
self.model = Cohere2MoeModel(
|
|
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
|
)
|
|
self.make_empty_intermediate_tensors = (
|
|
self.model.make_empty_intermediate_tensors
|
|
)
|
|
|
|
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.model.get_input_embeddings(input_ids)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.model.get_input_embeddings(input_ids)
|
|
|
|
@torch.no_grad()
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None = None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
) -> torch.Tensor | IntermediateTensors:
|
|
return self.model(input_ids, positions, intermediate_tensors, inputs_embeds)
|
|
|
|
def compute_logits(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
) -> torch.Tensor | None:
|
|
return self.logits_processor(self.model.embed_tokens, hidden_states)
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
loader = AutoWeightsLoader(self, skip_prefixes=["lm_head."])
|
|
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|