457 lines
16 KiB
Python
457 lines
16 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
# Adapted from
|
|
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
|
|
# Copyright 2023 The vLLM team.
|
|
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
|
# and OPT implementations in this library. It has been modified from its
|
|
# original forms to accommodate minor architectural differences compared
|
|
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
"""Inference-only GraniteMoe model."""
|
|
|
|
from collections.abc import Iterable
|
|
from itertools import islice
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers.configuration_utils import PretrainedConfig
|
|
|
|
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,
|
|
tensor_model_parallel_all_gather,
|
|
)
|
|
from vllm.model_executor.layers.attention import Attention
|
|
from vllm.model_executor.layers.fused_moe import FusedMoEFactory
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
|
from vllm.model_executor.layers.linear import (
|
|
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 (
|
|
ParallelLMHead,
|
|
VocabParallelEmbedding,
|
|
)
|
|
from vllm.model_executor.model_loader.weight_utils import sharded_weight_loader
|
|
from vllm.model_executor.models.utils import sequence_parallel_chunk
|
|
from vllm.model_executor.utils import set_weight_attrs
|
|
from vllm.sequence import IntermediateTensors
|
|
|
|
from .granite import granite_layer_attn_params
|
|
from .interfaces import SupportsLoRA, SupportsPP
|
|
from .utils import (
|
|
AutoWeightsLoader,
|
|
WeightsMapper,
|
|
extract_layer_index,
|
|
make_layers,
|
|
maybe_prefix,
|
|
)
|
|
|
|
|
|
class GraniteMoeMoE(nn.Module):
|
|
"""A tensor-parallel MoE implementation for GraniteMoe that shards each
|
|
expert across all ranks.
|
|
Each expert's weights are sharded across all ranks and a fused MoE
|
|
kernel is used for the forward pass, and finally we reduce the outputs
|
|
across ranks.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
num_experts: int,
|
|
top_k: int,
|
|
hidden_size: int,
|
|
intermediate_size: int,
|
|
params_dtype: torch.dtype | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
tp_size: int | None = None,
|
|
is_sequence_parallel=False,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
self.is_sequence_parallel = is_sequence_parallel
|
|
|
|
# Gate always runs at half / full precision for now.
|
|
self.gate = ReplicatedLinear(
|
|
hidden_size,
|
|
num_experts,
|
|
bias=False,
|
|
params_dtype=params_dtype,
|
|
quant_config=None,
|
|
prefix=f"{prefix}.gate",
|
|
)
|
|
|
|
self.experts = FusedMoEFactory(
|
|
num_experts=num_experts,
|
|
top_k=top_k,
|
|
hidden_size=hidden_size,
|
|
intermediate_size=intermediate_size,
|
|
params_dtype=params_dtype,
|
|
renormalize=True,
|
|
quant_config=quant_config,
|
|
tp_size=tp_size,
|
|
prefix=f"{prefix}.experts",
|
|
is_sequence_parallel=self.is_sequence_parallel,
|
|
)
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
# NOTE: hidden_states can have either 1D or 2D shape.
|
|
orig_shape = hidden_states.shape
|
|
hidden_states = hidden_states.view(-1, self.hidden_size)
|
|
|
|
if self.is_sequence_parallel:
|
|
hidden_states = sequence_parallel_chunk(hidden_states)
|
|
|
|
# router_logits: (num_tokens, n_experts)
|
|
router_logits, _ = self.gate(hidden_states)
|
|
final_hidden_states = self.experts(hidden_states, router_logits)
|
|
|
|
if self.is_sequence_parallel:
|
|
final_hidden_states = tensor_model_parallel_all_gather(
|
|
final_hidden_states, 0
|
|
)
|
|
num_tokens = orig_shape[0]
|
|
final_hidden_states = final_hidden_states[:num_tokens]
|
|
|
|
return final_hidden_states.view(orig_shape)
|
|
|
|
|
|
class GraniteMoeAttention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
max_position: int = 4096 * 32,
|
|
cache_config: CacheConfig | None = None,
|
|
quant_config: QuantizationConfig | None = None,
|
|
attention_multiplier: float | None = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
self.total_num_heads = num_heads
|
|
assert self.total_num_heads % tp_size == 0
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.total_num_kv_heads = num_kv_heads
|
|
if self.total_num_kv_heads >= tp_size:
|
|
# Number of KV heads is greater than TP size, so we partition
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert self.total_num_kv_heads % tp_size == 0
|
|
else:
|
|
# Number of KV heads is less than TP size, so we replicate
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
|
self.head_dim = hidden_size // self.total_num_heads
|
|
self.q_size = self.num_heads * self.head_dim
|
|
self.kv_size = self.num_kv_heads * self.head_dim
|
|
self.scaling = (
|
|
attention_multiplier
|
|
if attention_multiplier is not None
|
|
else self.head_dim**-1
|
|
)
|
|
|
|
self.qkv_proj = QKVParallelLinear(
|
|
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,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.o_proj",
|
|
)
|
|
sliding_window, rope_theta, has_sink = granite_layer_attn_params(
|
|
config, extract_layer_index(prefix)
|
|
)
|
|
|
|
self.use_rope = rope_theta != 0
|
|
if self.use_rope:
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
max_position=max_position,
|
|
rope_parameters={**config.rope_parameters, "rope_theta": rope_theta},
|
|
is_neox_style=True,
|
|
)
|
|
|
|
if has_sink:
|
|
self.sinks = nn.Parameter(torch.empty(self.num_heads), requires_grad=False)
|
|
set_weight_attrs(self.sinks, {"weight_loader": sharded_weight_loader(0)})
|
|
else:
|
|
self.sinks = None
|
|
|
|
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=sliding_window,
|
|
prefix=f"{prefix}.attn",
|
|
sinks=self.sinks,
|
|
)
|
|
|
|
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.use_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 GraniteMoeDecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
vllm_config: VllmConfig,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
config = vllm_config.model_config.hf_config
|
|
cache_config = vllm_config.cache_config
|
|
quant_config = vllm_config.quant_config
|
|
parallel_config = vllm_config.parallel_config
|
|
|
|
self.hidden_size = config.hidden_size
|
|
self.self_attn = GraniteMoeAttention(
|
|
config=config,
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
max_position=config.max_position_embeddings,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.self_attn",
|
|
attention_multiplier=config.attention_multiplier,
|
|
)
|
|
self.block_sparse_moe = GraniteMoeMoE(
|
|
num_experts=config.num_local_experts,
|
|
top_k=config.num_experts_per_tok,
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
quant_config=quant_config,
|
|
is_sequence_parallel=parallel_config.use_sequence_parallel_moe,
|
|
prefix=f"{prefix}.block_sparse_moe",
|
|
)
|
|
|
|
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.post_attention_layernorm = RMSNorm(
|
|
config.hidden_size, eps=config.rms_norm_eps
|
|
)
|
|
|
|
self.residual_multiplier = config.residual_multiplier
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
# Self Attention
|
|
residual = hidden_states
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
)
|
|
hidden_states = residual + hidden_states * self.residual_multiplier
|
|
residual = hidden_states
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
|
hidden_states = self.block_sparse_moe(hidden_states)
|
|
hidden_states = residual + hidden_states * self.residual_multiplier
|
|
|
|
return hidden_states
|
|
|
|
|
|
@support_torch_compile
|
|
class GraniteMoeModel(nn.Module):
|
|
hf_to_vllm_mapper: WeightsMapper = WeightsMapper(
|
|
orig_to_new_suffix={
|
|
# Legacy names to new names
|
|
"moe.input_linear.weight": "moe.experts.gate_up_proj",
|
|
"moe.output_linear.weight": "moe.experts.down_proj",
|
|
".router.layer.weight": ".gate.weight",
|
|
# Checkpoint name to vLLM name
|
|
".router.weight": ".gate.weight",
|
|
},
|
|
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"),
|
|
},
|
|
)
|
|
|
|
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
|
|
self.quant_config = quant_config # Required by MixtralModel
|
|
|
|
self.vocab_size = config.vocab_size
|
|
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
self.vocab_size,
|
|
config.hidden_size,
|
|
)
|
|
self.embedding_multiplier = config.embedding_multiplier
|
|
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda prefix: GraniteMoeDecoderLayer(vllm_config, prefix=prefix),
|
|
prefix=f"{prefix}.layers",
|
|
)
|
|
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.embed_tokens(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor | None,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
if get_pp_group().is_first_rank:
|
|
if inputs_embeds is not None:
|
|
hidden_states = inputs_embeds
|
|
else:
|
|
hidden_states = self.embed_input_ids(input_ids)
|
|
hidden_states *= self.embedding_multiplier
|
|
else:
|
|
assert intermediate_tensors is not None
|
|
hidden_states = intermediate_tensors["hidden_states"]
|
|
for layer in islice(self.layers, self.start_layer, self.end_layer):
|
|
hidden_states = layer(positions, hidden_states)
|
|
if not get_pp_group().is_last_rank:
|
|
return IntermediateTensors(
|
|
{
|
|
"hidden_states": hidden_states,
|
|
}
|
|
)
|
|
hidden_states = self.norm(hidden_states)
|
|
return hidden_states
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
|
|
|
|
|
class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
|
fall_back_to_pt_during_load = False
|
|
|
|
packed_modules_mapping = {
|
|
"qkv_proj": [
|
|
"q_proj",
|
|
"k_proj",
|
|
"v_proj",
|
|
],
|
|
}
|
|
|
|
# LoRA specific attributes
|
|
embedding_modules = {
|
|
"embed_tokens": "input_embeddings",
|
|
"lm_head": "output_embeddings",
|
|
}
|
|
|
|
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
|
|
|
|
self.model = GraniteMoeModel(
|
|
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
|
)
|
|
|
|
self.lm_head = ParallelLMHead(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
prefix=maybe_prefix(prefix, "lm_head"),
|
|
)
|
|
if config.tie_word_embeddings:
|
|
self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
|
|
|
|
self.logits_processor = LogitsProcessor(
|
|
config.vocab_size,
|
|
scale=1 / self.config.logits_scaling,
|
|
)
|
|
|
|
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.model.embed_input_ids(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor | None,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: IntermediateTensors | None = None,
|
|
inputs_embeds: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(
|
|
input_ids, positions, intermediate_tensors, inputs_embeds
|
|
)
|
|
return hidden_states
|
|
|
|
def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
|
|
logits = self.logits_processor(self.lm_head, hidden_states)
|
|
return logits
|
|
|
|
def make_empty_intermediate_tensors(
|
|
self, batch_size: int, dtype: torch.dtype, device: torch.device
|
|
) -> IntermediateTensors:
|
|
return IntermediateTensors(
|
|
{
|
|
"hidden_states": torch.zeros(
|
|
(batch_size, self.config.hidden_size), dtype=dtype, device=device
|
|
),
|
|
}
|
|
)
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(weights)
|