biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

202
vllm/model_executor/models/commandr.py
View File

@@ -22,6 +22,7 @@
# This file is based on the LLama model definition file in transformers
"""PyTorch Cohere model."""
from collections.abc import Iterable
from itertools import islice
from typing import Optional, Union
@@ -35,26 +36,33 @@ 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.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
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 (
VocabParallelEmbedding)
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader, maybe_remap_kv_scale_name,
row_parallel_weight_loader)
default_weight_loader,
maybe_remap_kv_scale_name,
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 .interfaces import SupportsLoRA, SupportsPP, SupportsQuant
from .utils import (AutoWeightsLoader, extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
@torch.compile(backend=current_platform.simple_compile_backend)
@@ -63,30 +71,27 @@ def layer_norm_func(hidden_states, weight, variance_epsilon):
hidden_states = hidden_states.to(torch.float32)
mean = hidden_states.mean(-1, keepdim=True)
variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
hidden_states = (hidden_states - mean) * torch.rsqrt(variance +
variance_epsilon)
hidden_states = (hidden_states - mean) * torch.rsqrt(variance + variance_epsilon)
hidden_states = weight.to(torch.float32) * hidden_states
return hidden_states.to(input_dtype)
class LayerNorm(nn.Module):
def __init__(self, param_shape=None, eps=1e-5):
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})
set_weight_attrs(self.weight, {"weight_loader": row_parallel_weight_loader})
def forward(self, hidden_states, residuals=None):
hidden_states = layer_norm_func(hidden_states, self.weight,
self.variance_epsilon)
hidden_states = layer_norm_func(
hidden_states, self.weight, self.variance_epsilon
)
return hidden_states, residuals
# Copied from transformers.models.llama.modeling_llama.LlamaMLP Llama->Cohere
class CohereMLP(nn.Module):
def __init__(
self,
config: Union[CohereConfig, Cohere2Config],
@@ -121,7 +126,6 @@ class CohereMLP(nn.Module):
class CohereAttention(nn.Module):
def __init__(
self,
config: Union[CohereConfig, Cohere2Config],
@@ -151,8 +155,8 @@ class CohereAttention(nn.Module):
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)
config, "model_max_length", None
) or getattr(config, "max_position_embeddings", 8192)
self.rope_theta = config.rope_theta
self.rope_scaling = getattr(config, "rope_scaling", None)
self.use_qk_norm = getattr(config, "use_qk_norm", False)
@@ -190,21 +194,24 @@ class CohereAttention(nn.Module):
if config.layer_types[layer_idx] == "sliding_attention":
self.sliding_window = config.sliding_window
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")
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",
)
if self.use_qk_norm:
self.q_norm = LayerNorm(param_shape=(self.num_heads,
self.head_dim),
eps=config.layer_norm_eps)
self.k_norm = LayerNorm(param_shape=(self.num_kv_heads,
self.head_dim),
eps=config.layer_norm_eps)
self.q_norm = LayerNorm(
param_shape=(self.num_heads, self.head_dim), eps=config.layer_norm_eps
)
self.k_norm = LayerNorm(
param_shape=(self.num_kv_heads, self.head_dim),
eps=config.layer_norm_eps,
)
def _apply_qk_norm(self, q, k):
q = q.view(*q.shape[:-1], -1, self.head_dim)
@@ -232,25 +239,27 @@ class CohereAttention(nn.Module):
class CohereDecoderLayer(nn.Module):
def __init__(self,
config: Union[CohereConfig, Cohere2Config],
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = ""):
def __init__(
self,
config: Union[CohereConfig, Cohere2Config],
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = CohereAttention(config,
cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn")
self.self_attn = CohereAttention(
config,
cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
)
self.mlp = CohereMLP(config,
quant_config=quant_config,
prefix=f"{prefix}.mlp")
self.input_layernorm = LayerNorm(param_shape=(config.hidden_size),
eps=config.layer_norm_eps)
self.mlp = CohereMLP(config, quant_config=quant_config, prefix=f"{prefix}.mlp")
self.input_layernorm = LayerNorm(
param_shape=(config.hidden_size), eps=config.layer_norm_eps
)
def forward(
self,
@@ -274,7 +283,6 @@ class CohereDecoderLayer(nn.Module):
@support_torch_compile
class CohereModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -285,22 +293,29 @@ class CohereModel(nn.Module):
self.quant_config = quant_config
self.config = config
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
lora_vocab = (
(lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
if lora_config
else 0
)
self.vocab_size = config.vocab_size + lora_vocab
self.org_vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
config.hidden_size)
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size, config.hidden_size
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: CohereDecoderLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
self.norm = LayerNorm(param_shape=(config.hidden_size),
eps=config.layer_norm_eps)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
self.norm = LayerNorm(
param_shape=(config.hidden_size), eps=config.layer_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)
@@ -329,15 +344,13 @@ class CohereModel(nn.Module):
residual,
)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states": hidden_states,
"residual": residual
})
return IntermediateTensors(
{"hidden_states": hidden_states, "residual": residual}
)
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
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_proj", "q_proj", "q"),
@@ -349,14 +362,15 @@ class CohereModel(nn.Module):
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
if (self.quant_config is not None and
(scale_name := self.quant_config.get_cache_scale(name))):
if self.quant_config is not None and (
scale_name := self.quant_config.get_cache_scale(name)
):
# Loading kv cache quantization scales
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
loaded_weight[0])
weight_loader = getattr(param, "weight_loader", default_weight_loader)
loaded_weight = (
loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
)
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue
@@ -386,8 +400,7 @@ class CohereModel(nn.Module):
if is_pp_missing_parameter(name, self):
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, loaded_weight)
loaded_params.add(name)
return loaded_params
@@ -421,13 +434,15 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.quant_config = quant_config
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size,
scale=config.logit_scale)
self.model = CohereModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, scale=config.logit_scale
)
self.model = CohereModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
self.model.make_empty_intermediate_tensors
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
@@ -440,26 +455,27 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
hidden_states = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
def compute_logits(
self,
hidden_states: torch.Tensor,
) -> Optional[torch.Tensor]:
is_not_lora = hasattr(self.model.embed_tokens, 'weight')
is_not_lora = hasattr(self.model.embed_tokens, "weight")
if is_not_lora:
logits = self.logits_processor(self.model.embed_tokens,
hidden_states)
logits = self.logits_processor(self.model.embed_tokens, hidden_states)
else:
logits = self.logits_processor(self.model.embed_tokens.base_layer,
hidden_states)
logits = self.logits_processor(
self.model.embed_tokens.base_layer, 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", "rotary_emb.inv_freq"])
self, skip_prefixes=["lm_head", "rotary_emb.inv_freq"]
)
return loader.load_weights(weights)