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:
Harry Mellor
2025-10-05 07:06:22 -07:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115205 additions and 94107 deletions
+123 -96
View File
@@ -23,6 +23,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Nemotron model compatible with HuggingFace weights."""
from collections.abc import Iterable
from itertools import islice
from typing import Any, Optional, Union
@@ -35,23 +36,35 @@ 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 get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
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 (
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader, maybe_remap_kv_scale_name)
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import NemotronConfig
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
# The architecture is pretty similar to Llama, with these changes:
# - There is no gate_proj, just up_proj
@@ -65,20 +78,21 @@ def _cast_if_autocast_enabled(*args):
return args
else:
return torch.amp.autocast_mode._cast(
args, device_type="cuda", dtype=torch.get_autocast_gpu_dtype())
args, device_type="cuda", dtype=torch.get_autocast_gpu_dtype()
)
class NemotronLayerNorm1P(nn.LayerNorm):
def __init__(self,
normalized_shape: Union[int, list[int], torch.Size],
eps: float = 1e-5,
elementwise_affine: bool = True,
bias: bool = True,
device=None,
dtype=None):
super().__init__(normalized_shape, eps, elementwise_affine, bias,
device, dtype)
def __init__(
self,
normalized_shape: Union[int, list[int], torch.Size],
eps: float = 1e-5,
elementwise_affine: bool = True,
bias: bool = True,
device=None,
dtype=None,
):
super().__init__(normalized_shape, eps, elementwise_affine, bias, device, dtype)
def forward(
self,
@@ -88,15 +102,15 @@ class NemotronLayerNorm1P(nn.LayerNorm):
if residual is not None:
x = x + residual
residual = x
args = _cast_if_autocast_enabled(x, self.normalized_shape,
self.weight + 1, self.bias, self.eps)
args = _cast_if_autocast_enabled(
x, self.normalized_shape, self.weight + 1, self.bias, self.eps
)
with torch.amp.autocast("cuda", enabled=False):
x = torch.nn.functional.layer_norm(*args)
return x if residual is None else (x, residual)
class NemotronMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -107,16 +121,20 @@ class NemotronMLP(nn.Module):
prefix: str = "",
) -> None:
super().__init__()
self.up_proj = ColumnParallelLinear(input_size=hidden_size,
output_size=intermediate_size,
bias=bias,
quant_config=quant_config,
prefix=f"{prefix}.up_proj")
self.down_proj = RowParallelLinear(input_size=intermediate_size,
output_size=hidden_size,
bias=bias,
quant_config=quant_config,
prefix=f"{prefix}.down_proj")
self.up_proj = ColumnParallelLinear(
input_size=hidden_size,
output_size=intermediate_size,
bias=bias,
quant_config=quant_config,
prefix=f"{prefix}.up_proj",
)
self.down_proj = RowParallelLinear(
input_size=intermediate_size,
output_size=hidden_size,
bias=bias,
quant_config=quant_config,
prefix=f"{prefix}.down_proj",
)
self.act_fn = get_act_fn(hidden_act)
def forward(self, x):
@@ -127,7 +145,6 @@ class NemotronMLP(nn.Module):
class NemotronAttention(nn.Module):
def __init__(
self,
config: NemotronConfig,
@@ -194,13 +211,15 @@ class NemotronAttention(nn.Module):
rope_scaling=rope_scaling,
partial_rotary_factor=self.partial_rotary_factor,
)
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,
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,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -216,7 +235,6 @@ class NemotronAttention(nn.Module):
class NemotronDecoderLayer(nn.Module):
def __init__(
self,
config: NemotronConfig,
@@ -229,21 +247,24 @@ class NemotronDecoderLayer(nn.Module):
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None):
config, "original_max_position_embeddings", None
):
rope_scaling["original_max_position_embeddings"] = (
config.original_max_position_embeddings)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
config.original_max_position_embeddings
)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
# Support abacusai/Smaug-72B-v0.1 with attention_bias
# Support internlm/internlm-7b with bias
attention_bias = getattr(config, "attention_bias", False) or getattr(
config, "bias", False)
config, "bias", False
)
self.self_attn = NemotronAttention(
config=config,
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
num_kv_heads=getattr(config, "num_key_value_heads",
config.num_attention_heads),
num_kv_heads=getattr(
config, "num_key_value_heads", config.num_attention_heads
),
rope_theta=rope_theta,
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
@@ -260,10 +281,12 @@ class NemotronDecoderLayer(nn.Module):
bias=getattr(config, "mlp_bias", False),
prefix=f"{prefix}.mlp",
)
self.input_layernorm = NemotronLayerNorm1P(config.hidden_size,
eps=config.norm_eps)
self.input_layernorm = NemotronLayerNorm1P(
config.hidden_size, eps=config.norm_eps
)
self.post_attention_layernorm = NemotronLayerNorm1P(
config.hidden_size, eps=config.norm_eps)
config.hidden_size, eps=config.norm_eps
)
def forward(
self,
@@ -276,23 +299,20 @@ class NemotronDecoderLayer(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.self_attn(
positions=positions,
hidden_states=hidden_states,
)
# Fully Connected
hidden_states, residual = self.post_attention_layernorm(
hidden_states, residual)
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
hidden_states = self.mlp(hidden_states)
return hidden_states, residual
@support_torch_compile
class NemotronModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -303,12 +323,16 @@ class NemotronModel(nn.Module):
self.config = config
self.quant_config = quant_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
if get_pp_group().is_first_rank or (config.tie_word_embeddings
and get_pp_group().is_last_rank):
if get_pp_group().is_first_rank or (
config.tie_word_embeddings and get_pp_group().is_last_rank
):
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
@@ -318,19 +342,21 @@ class NemotronModel(nn.Module):
self.embed_tokens = PPMissingLayer()
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: NemotronDecoderLayer(config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix),
prefix=f"{prefix}.layers")
lambda prefix: NemotronDecoderLayer(
config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix,
),
prefix=f"{prefix}.layers",
)
if get_pp_group().is_last_rank:
self.norm = NemotronLayerNorm1P(config.hidden_size,
eps=config.norm_eps)
self.norm = NemotronLayerNorm1P(config.hidden_size, eps=config.norm_eps)
else:
self.norm = PPMissingLayer()
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
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)
@@ -357,16 +383,14 @@ class NemotronModel(nn.Module):
hidden_states, residual = layer(positions, hidden_states, 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"),
@@ -376,18 +400,19 @@ class NemotronModel(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
for (param_name, weight_name, shard_id) in stacked_params_mapping:
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
@@ -416,8 +441,7 @@ class NemotronModel(nn.Module):
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
@@ -450,8 +474,9 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.lora_config = lora_config
self.quant_config = quant_config
self.model = NemotronModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = NemotronModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
if get_pp_group().is_last_rank:
self.unpadded_vocab_size = config.vocab_size
if lora_config:
@@ -463,7 +488,8 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
if not lora_config
else lora_config.lora_vocab_padding_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
@@ -471,14 +497,15 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.lm_head.weight = self.model.embed_tokens.weight
logit_scale = getattr(config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size,
logit_scale)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, logit_scale
)
else:
self.lm_head = PPMissingLayer()
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)
@@ -490,8 +517,9 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
model_output = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
model_output = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return model_output
def compute_logits(
@@ -501,7 +529,6 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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)
return loader.load_weights(weights)