Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
769 lines
25 KiB
Python
769 lines
25 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Based on:
|
|
Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023).
|
|
Punica: Multi-Tenant LoRA Serving.
|
|
https://arxiv.org/abs/2310.18547
|
|
"""
|
|
|
|
from typing import Any, final
|
|
|
|
import torch
|
|
|
|
from vllm.lora.layers import LoRAMapping
|
|
from vllm.lora.utils import get_captured_lora_counts
|
|
from vllm.triton_utils import HAS_TRITON, triton
|
|
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
|
|
from vllm.utils.math_utils import round_up
|
|
|
|
if HAS_TRITON:
|
|
from vllm.lora.ops.triton_ops import (
|
|
LoRAKernelMeta,
|
|
fused_moe_lora,
|
|
lora_expand,
|
|
lora_shrink,
|
|
)
|
|
|
|
from vllm import _custom_ops as ops
|
|
|
|
from .punica_base import PunicaWrapperBase
|
|
|
|
|
|
@final
|
|
class PunicaWrapperGPU(PunicaWrapperBase):
|
|
"""PunicaWrapperGPU is designed to manage and provide metadata for the punica
|
|
kernel. The main function is to maintain the state information for
|
|
Multi-LoRA, and to provide the interface for the punica triton kernel.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
max_num_batched_tokens: int,
|
|
max_batches: int,
|
|
device: torch.device | str,
|
|
**kwargs,
|
|
):
|
|
PunicaWrapperBase.__init__(self, max_num_batched_tokens, max_batches, device)
|
|
|
|
self.lora_config = kwargs["lora_config"]
|
|
self.max_loras = self.lora_config.max_loras
|
|
|
|
# Compute captured LoRA counts for cudagraph specialization.
|
|
captured_lora_counts = get_captured_lora_counts(
|
|
self.max_loras, self.lora_config.specialize_active_lora
|
|
)
|
|
|
|
self.token_mapping_meta = LoRAKernelMeta.make(
|
|
self.max_loras,
|
|
max_num_batched_tokens,
|
|
device=device,
|
|
captured_lora_counts=captured_lora_counts,
|
|
)
|
|
|
|
# When speculative decoding is enabled, max_num_samples is
|
|
# max_batches * (num_speculative_decoding_tokens + 1).
|
|
# This line can be optimized by replacing max_num_batched_tokens
|
|
# to max_batches * (num_speculative_decoding_tokens + 1).
|
|
self.prompt_mapping_meta = LoRAKernelMeta.make(
|
|
self.max_loras,
|
|
max_num_batched_tokens,
|
|
device=device,
|
|
captured_lora_counts=captured_lora_counts,
|
|
)
|
|
|
|
def update_metadata(
|
|
self,
|
|
mapping: LoRAMapping,
|
|
lora_index_to_id: list[int | None],
|
|
max_loras: int,
|
|
vocab_size: int,
|
|
**kwargs,
|
|
):
|
|
self.is_prefill = mapping.is_prefill
|
|
self._update_base_metadata(mapping, lora_index_to_id, max_loras, vocab_size)
|
|
|
|
# TODO avoid gpu<->cpu sync here
|
|
with gpu_sync_allowed():
|
|
# Prepare cuda kernel metadata tensors
|
|
self.token_mapping_meta.prepare_tensors(self.token_lora_indices)
|
|
self.prompt_mapping_meta.prepare_tensors(self.sampler_indices)
|
|
|
|
def add_shrink(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: tuple[torch.Tensor, ...],
|
|
scale: float,
|
|
**kwargs: Any,
|
|
):
|
|
"""Performs GEMM for multiple slices of lora_a.
|
|
|
|
Semantics:
|
|
for i in range(len(lora_a_stacked)):
|
|
y[i] += (x @ lora_a_stacked[i]) * scale
|
|
|
|
Args:
|
|
y (torch.Tensor): Output tensors
|
|
x (torch.Tensor): Input tensor
|
|
lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weights
|
|
scale (float): Scaling factor for the operation
|
|
**kwargs: Unused; accepted for compatibility with the base class signature.
|
|
|
|
"""
|
|
x = x.view(-1, x.shape[-1])
|
|
lora_shrink(
|
|
x,
|
|
lora_a_stacked,
|
|
y,
|
|
*self.token_mapping_meta.meta_args(
|
|
x.size(0), self.lora_config.specialize_active_lora
|
|
),
|
|
scale,
|
|
)
|
|
|
|
def add_expand(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_b_stacked: tuple[torch.Tensor, ...],
|
|
output_slices: tuple[int, ...],
|
|
offset_start: int = 0,
|
|
add_inputs=True,
|
|
**kwargs: Any,
|
|
) -> None:
|
|
"""Performs GEMM for multiple slices of lora_b.
|
|
|
|
Semantics:
|
|
for i in range(len(lora_b_stacked)):
|
|
slice = output_slices[i]
|
|
y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i]
|
|
offset += slice
|
|
|
|
Args:
|
|
y (torch.Tensor): Output tensor.
|
|
x (torch.Tensor): Input tensors
|
|
lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight
|
|
output_slices (tuple[int, ...]): Every slice's size
|
|
offset_start (int): The starting position of y, defaults to 0.
|
|
add_inputs (bool): If True, add LoRA output to y; if False, write
|
|
LoRA-only output to y (used for dual-stream when base and LoRA
|
|
run on different CUDA streams). Defaults to True.
|
|
**kwargs: Unused; accepted for compatibility with the base class signature.
|
|
|
|
"""
|
|
y_org = y
|
|
y = y.view(-1, y.shape[-1])
|
|
|
|
assert x.ndim == 3
|
|
assert x.size(0) == len(output_slices)
|
|
num_tokens = x.size(1) # first dimension is the num slices
|
|
|
|
lora_expand(
|
|
x,
|
|
lora_b_stacked,
|
|
y,
|
|
*self.token_mapping_meta.meta_args(
|
|
num_tokens, self.lora_config.specialize_active_lora
|
|
),
|
|
offset_start=offset_start,
|
|
add_inputs=add_inputs,
|
|
)
|
|
|
|
y = y.view_as(y_org)
|
|
|
|
def add_lora_embedding(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_b_stacked: torch.Tensor,
|
|
add_inputs: bool = True,
|
|
**kwargs: Any,
|
|
) -> None:
|
|
"""Applies lora specifically for VocabParallelEmbeddingWithLoRA.
|
|
|
|
Semantics:
|
|
y += x @ lora_b_stacked
|
|
|
|
Args:
|
|
y (torch.Tensor): Output tensor.
|
|
x (torch.Tensor): Input tensor.
|
|
lora_b_stacked (torch.Tensor): lora_b's weights.
|
|
add_inputs (bool): Default to True.
|
|
**kwargs: Unused; accepted for compatibility with the base class signature.
|
|
|
|
"""
|
|
lora_expand(
|
|
x.unsqueeze(dim=0),
|
|
(lora_b_stacked,),
|
|
y,
|
|
*self.token_mapping_meta.meta_args(
|
|
x.size(0), self.lora_config.specialize_active_lora
|
|
),
|
|
offset_start=0,
|
|
add_inputs=add_inputs,
|
|
)
|
|
|
|
def add_lora_linear(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: tuple[torch.Tensor, ...],
|
|
lora_b_stacked: tuple[torch.Tensor, ...],
|
|
scale: float,
|
|
output_slices: tuple[int, ...],
|
|
*,
|
|
buffer: torch.Tensor | None = None,
|
|
**kwargs: Any,
|
|
) -> None:
|
|
"""Applicable to linear-related lora.
|
|
|
|
Semantics:
|
|
for i in range(len(lora_a_stacked)):
|
|
y[i] += (
|
|
x[i].unsqueeze(0)
|
|
@ lora_a_stacked[indices[i], layer_idx, :, :]
|
|
@ lora_b_stacked[indices[i], layer_idx, :, :]
|
|
* scale
|
|
).squeeze(0)
|
|
|
|
Args:
|
|
y (torch.Tensor): Output tensor. Will be changed in-place.
|
|
x (torch.Tensor): Input tensor
|
|
lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weight.
|
|
lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight.
|
|
scale (float): Scaling factor.
|
|
output_slices (tuple[int, ...]): Every slice's size.
|
|
buffer (Optional[torch.Tensor]): Defaults to None.
|
|
**kwargs: Accepts `add_inputs` (bool, defaults to True); any other
|
|
keys are ignored.
|
|
|
|
"""
|
|
assert len(lora_a_stacked) == len(lora_b_stacked) == len(output_slices)
|
|
|
|
assert buffer is None, (
|
|
"To minimize overhead, the buffer should be created by "
|
|
".add_lora_linear() instead of being passed in."
|
|
)
|
|
r = lora_b_stacked[0].size(-1)
|
|
# We set the buffer to be float32 by default, refer to:
|
|
# https://github.com/triton-lang/triton/issues/1387
|
|
# Note: buffer is zeroed inside the shrink op
|
|
buffer = torch.empty(
|
|
(len(output_slices), x.size(0), r), dtype=torch.float32, device=x.device
|
|
)
|
|
add_inputs = kwargs.pop("add_inputs", True)
|
|
self.add_shrink(
|
|
buffer, # type: ignore
|
|
x,
|
|
lora_a_stacked,
|
|
scale,
|
|
**kwargs,
|
|
)
|
|
self.add_expand(
|
|
y,
|
|
buffer, # type: ignore
|
|
lora_b_stacked,
|
|
output_slices,
|
|
add_inputs=add_inputs,
|
|
**kwargs,
|
|
)
|
|
|
|
def add_lora_logits(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: torch.Tensor,
|
|
lora_b_stacked: torch.Tensor,
|
|
scale,
|
|
*,
|
|
buffer: torch.Tensor | None = None,
|
|
**kwargs: Any,
|
|
) -> None:
|
|
"""Applies lora specifically for LogitsProcessorWithLoRA.
|
|
|
|
Semantics:
|
|
buffer = (x @ lora_a_stacked) * scale
|
|
y += buffer @ lora_b_stacked
|
|
|
|
Args:
|
|
y (torch.Tensor): Output tensor.
|
|
x (torch.Tensor): Input tensor.
|
|
lora_a_stacked (torch.Tensor): lora_a's weights.
|
|
lora_b_stacked (torch.Tensor): lora_b's weights.
|
|
scale (float): Scaling factor.
|
|
buffer (Optional[torch.Tensor]): Default to None.
|
|
**kwargs: Unused; accepted for compatibility with the base class signature.
|
|
|
|
"""
|
|
y_org = y
|
|
y = y.view(-1, y.shape[-1])
|
|
x = x.view(-1, x.shape[-1])
|
|
r = lora_b_stacked.size(-1)
|
|
|
|
assert buffer is None, (
|
|
"To minimize overhead, the buffer should be created by "
|
|
".add_lora_linear() instead of being passed in."
|
|
)
|
|
# We set the buffer to be float32 by default, refer to:
|
|
# https://github.com/triton-lang/triton/issues/1387
|
|
# Note: buffer is zeroed inside the shrink op
|
|
buffer = torch.empty((x.size(0), r), dtype=torch.float32, device=x.device)
|
|
|
|
lora_shrink(
|
|
x,
|
|
[lora_a_stacked],
|
|
buffer.unsqueeze(dim=0),
|
|
*self.prompt_mapping_meta.meta_args(
|
|
x.size(0), self.lora_config.specialize_active_lora
|
|
),
|
|
scale,
|
|
)
|
|
|
|
lora_expand(
|
|
buffer.unsqueeze(dim=0),
|
|
[lora_b_stacked],
|
|
y,
|
|
*self.prompt_mapping_meta.meta_args(
|
|
buffer.size(0), self.lora_config.specialize_active_lora
|
|
),
|
|
add_inputs=True,
|
|
)
|
|
y = y.view_as(y_org)
|
|
|
|
def apply_lora_full_linear(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
weight_stacked: torch.Tensor,
|
|
bias_stacked: torch.Tensor,
|
|
module_enabled: torch.Tensor,
|
|
) -> None:
|
|
adapter_y = torch.empty(
|
|
(x.size(0), weight_stacked.size(-2)),
|
|
dtype=torch.float32,
|
|
device=x.device,
|
|
)
|
|
lora_shrink(
|
|
x,
|
|
[weight_stacked],
|
|
adapter_y.unsqueeze(0),
|
|
*self.prompt_mapping_meta.meta_args(
|
|
x.size(0), self.lora_config.specialize_active_lora
|
|
),
|
|
1.0,
|
|
)
|
|
result = self._select_full_linear_output(
|
|
y, adapter_y, bias_stacked, module_enabled
|
|
)
|
|
y.copy_(result)
|
|
|
|
def moe_lora_align_block_size(
|
|
self,
|
|
topk_ids: torch.Tensor,
|
|
num_tokens: int,
|
|
block_size: int,
|
|
num_experts: int,
|
|
max_loras: int,
|
|
adapter_enabled: torch.Tensor,
|
|
expert_map: torch.Tensor | None = None,
|
|
pad_sorted_ids: bool = False,
|
|
naive_block_assignment: bool = False,
|
|
token_lora_mapping: torch.Tensor | None = None,
|
|
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
"""Aligns tokens and experts into block-sized chunks for LoRA-based
|
|
mixture-of-experts (MoE) execution.
|
|
|
|
When `token_lora_mapping` is provided, it overrides the global mapping
|
|
read from `self.token_mapping_meta`. This is how EP+LoRA injects the
|
|
per-rank-local token→LoRA map after all-to-all dispatch.
|
|
"""
|
|
(
|
|
token_lora_mapping_meta,
|
|
_,
|
|
_,
|
|
_,
|
|
lora_ids,
|
|
_,
|
|
_,
|
|
) = self.token_mapping_meta.meta_args(
|
|
num_tokens, self.lora_config.specialize_active_lora
|
|
)
|
|
if token_lora_mapping is None:
|
|
token_lora_mapping = token_lora_mapping_meta
|
|
# Under EP the caller passes local_num_experts but topk_ids carries
|
|
# GLOBAL expert indices. The CUDA kernel uses num_experts to size
|
|
# its bucketing table; with EP we must size by global_num_experts
|
|
# so global topk_ids don't overflow. expert_map inside the kernel
|
|
# then translates global→local so the output expert_ids are local
|
|
# (mirrors the non-LoRA moe_align_block_size behavior).
|
|
kernel_num_experts = (
|
|
expert_map.numel() if expert_map is not None else num_experts
|
|
)
|
|
if naive_block_assignment:
|
|
expert_ids = topk_ids.reshape(-1)
|
|
sorted_ids = None
|
|
num_tokens_post_pad = None
|
|
else:
|
|
max_num_tokens_padded = topk_ids.numel() + kernel_num_experts * (
|
|
block_size - 1
|
|
)
|
|
if pad_sorted_ids:
|
|
max_num_tokens_padded = round_up(max_num_tokens_padded, block_size)
|
|
if topk_ids.numel() < kernel_num_experts:
|
|
max_num_tokens_padded = topk_ids.numel() * block_size
|
|
sorted_ids = torch.empty(
|
|
(max_loras * max_num_tokens_padded,),
|
|
dtype=torch.int32,
|
|
device=topk_ids.device,
|
|
)
|
|
max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
|
|
# Expert ids are initialized to -1 so unused (lora, expert)
|
|
# slots don't drive the LoRA Triton kernel into the wrong bucket.
|
|
# The kernel overwrites only active slots.
|
|
expert_ids = torch.full(
|
|
(max_loras * max_num_m_blocks,),
|
|
-1,
|
|
dtype=torch.int32,
|
|
device=topk_ids.device,
|
|
)
|
|
num_tokens_post_pad = torch.empty(
|
|
(max_loras), dtype=torch.int32, device=topk_ids.device
|
|
)
|
|
|
|
ops.moe_lora_align_block_size(
|
|
topk_ids,
|
|
token_lora_mapping,
|
|
kernel_num_experts,
|
|
block_size,
|
|
max_loras,
|
|
max_num_tokens_padded,
|
|
max_num_m_blocks,
|
|
sorted_ids,
|
|
expert_ids,
|
|
num_tokens_post_pad,
|
|
adapter_enabled,
|
|
lora_ids,
|
|
expert_map,
|
|
)
|
|
|
|
return token_lora_mapping, sorted_ids, expert_ids, num_tokens_post_pad
|
|
|
|
def add_lora_fused_moe(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: tuple[torch.Tensor, ...],
|
|
lora_b_stacked: tuple[torch.Tensor, ...],
|
|
topk_weights: torch.Tensor,
|
|
sorted_token_ids: torch.Tensor | None,
|
|
expert_ids: torch.Tensor,
|
|
num_tokens_post_padded: torch.Tensor | None,
|
|
max_lora_rank: int,
|
|
top_k_num: int,
|
|
shrink_config,
|
|
expand_config,
|
|
adapter_enabled: torch.Tensor,
|
|
mul_routed_weight=False,
|
|
fully_sharded: bool = False,
|
|
offset: int = 0,
|
|
token_lora_mapping: torch.Tensor | None = None,
|
|
add_inputs: bool = True,
|
|
):
|
|
"""Perform a fused forward computation for a LoRA MoE layer."""
|
|
(
|
|
token_lora_mapping_meta,
|
|
_,
|
|
_,
|
|
_,
|
|
lora_ids,
|
|
no_lora_flag,
|
|
num_active_loras,
|
|
) = self.token_mapping_meta.meta_args(
|
|
x.size(0), self.lora_config.specialize_active_lora
|
|
)
|
|
|
|
assert no_lora_flag.numel() == 1
|
|
if no_lora_flag.item():
|
|
# None of the inputs require LoRA.
|
|
return
|
|
|
|
if token_lora_mapping is None:
|
|
token_lora_mapping = token_lora_mapping_meta
|
|
fused_moe_lora(
|
|
y,
|
|
x,
|
|
lora_a_stacked,
|
|
lora_b_stacked,
|
|
topk_weights,
|
|
sorted_token_ids,
|
|
expert_ids,
|
|
num_tokens_post_padded,
|
|
token_lora_mapping,
|
|
max_lora_rank,
|
|
top_k_num,
|
|
lora_ids,
|
|
num_active_loras,
|
|
adapter_enabled,
|
|
shrink_config.get("BLOCK_SIZE_M", 64),
|
|
shrink_config.get("BLOCK_SIZE_N", 64),
|
|
shrink_config.get("BLOCK_SIZE_K", 32),
|
|
shrink_config.get("GROUP_SIZE_M", 8),
|
|
shrink_config.get("NUM_WARPS", 4),
|
|
shrink_config.get("NUM_STAGES", 3),
|
|
shrink_config.get("SPLIT_K", 1),
|
|
expand_config.get("BLOCK_SIZE_M", 64),
|
|
expand_config.get("BLOCK_SIZE_N", 64),
|
|
expand_config.get("BLOCK_SIZE_K", 32),
|
|
expand_config.get("GROUP_SIZE_M", 8),
|
|
expand_config.get("NUM_WARPS", 4),
|
|
expand_config.get("NUM_STAGES", 3),
|
|
expand_config.get("SPLIT_K", 1),
|
|
mul_routed_weight,
|
|
fully_sharded,
|
|
offset,
|
|
add_inputs,
|
|
)
|
|
|
|
def add_lora_w13(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: tuple[torch.Tensor, ...],
|
|
lora_b_stacked: tuple[torch.Tensor, ...],
|
|
topk_ids: torch.Tensor,
|
|
topk_weights: torch.Tensor,
|
|
expert_map: torch.Tensor | None,
|
|
w1: torch.Tensor,
|
|
w2: torch.Tensor,
|
|
num_tokens: int,
|
|
top_k_num: int,
|
|
max_loras: int,
|
|
adapter_enabled: torch.Tensor,
|
|
local_num_experts: int,
|
|
top_k: int,
|
|
num_slices: int,
|
|
fully_sharded: bool,
|
|
use_tuned_config: bool,
|
|
add_inputs: bool = True,
|
|
token_lora_mapping: torch.Tensor | None = None,
|
|
) -> tuple[
|
|
torch.Tensor | None,
|
|
torch.Tensor | None,
|
|
torch.Tensor | None,
|
|
torch.Tensor | None,
|
|
]:
|
|
import functools
|
|
|
|
from vllm.lora.layers.utils import try_get_optimal_moe_lora_config
|
|
from vllm.lora.ops.triton_ops.utils import (
|
|
_normalize_lora_config_keys,
|
|
get_lora_op_configs,
|
|
)
|
|
from vllm.model_executor.layers.fused_moe.config import _get_config_dtype_str
|
|
|
|
config_dtype = _get_config_dtype_str(
|
|
dtype=x.dtype,
|
|
use_fp8_w8a8=False,
|
|
use_int8_w8a16=False,
|
|
use_int4_w4a16=False,
|
|
)
|
|
max_lora_rank = lora_a_stacked[0].shape[-2]
|
|
|
|
if use_tuned_config:
|
|
shrink_config = get_lora_op_configs(
|
|
op_type="fused_moe_lora_w13_shrink",
|
|
max_loras=max_loras,
|
|
batch=num_tokens,
|
|
hidden_size=x.shape[-1],
|
|
rank=max_lora_rank,
|
|
num_slices=num_slices,
|
|
moe_intermediate_size=lora_b_stacked[0].shape[-2],
|
|
)
|
|
expand_config = get_lora_op_configs(
|
|
op_type="fused_moe_lora_w13_expand",
|
|
max_loras=max_loras,
|
|
batch=num_tokens,
|
|
hidden_size=x.shape[-1],
|
|
rank=max_lora_rank,
|
|
num_slices=num_slices,
|
|
moe_intermediate_size=lora_b_stacked[0].shape[-2],
|
|
)
|
|
else:
|
|
get_config = functools.partial(
|
|
try_get_optimal_moe_lora_config,
|
|
w1_shape=w1.shape,
|
|
w2_shape=w2.shape,
|
|
rank=max_lora_rank,
|
|
top_k=top_k,
|
|
dtype=config_dtype,
|
|
M=num_tokens,
|
|
)
|
|
shrink_config = get_config(op_type="fused_moe_lora_w13_shrink")
|
|
expand_config = get_config(op_type="fused_moe_lora_w13_expand")
|
|
|
|
shrink_config = _normalize_lora_config_keys(shrink_config)
|
|
expand_config = _normalize_lora_config_keys(expand_config)
|
|
|
|
SPARSITY_FACTOR = 8
|
|
naive_block_assignment = (
|
|
not fully_sharded
|
|
and expert_map is None
|
|
and num_tokens * top_k * SPARSITY_FACTOR <= local_num_experts * max_loras
|
|
)
|
|
|
|
(
|
|
token_lora_mapping,
|
|
sorted_token_ids_lora,
|
|
expert_ids_lora,
|
|
num_tokens_post_padded_lora,
|
|
) = self.moe_lora_align_block_size(
|
|
topk_ids,
|
|
num_tokens,
|
|
int(shrink_config.get("BLOCK_SIZE_M") or 64),
|
|
local_num_experts,
|
|
max_loras,
|
|
adapter_enabled,
|
|
expert_map,
|
|
naive_block_assignment=naive_block_assignment,
|
|
token_lora_mapping=token_lora_mapping,
|
|
)
|
|
|
|
_sorted = sorted_token_ids_lora
|
|
_eids = expert_ids_lora
|
|
if _sorted is not None:
|
|
_eids = _eids.view(max_loras, -1)
|
|
_sorted = _sorted.view(max_loras, -1)
|
|
|
|
self.add_lora_fused_moe(
|
|
y.view(-1, top_k_num, y.shape[-1]),
|
|
x,
|
|
lora_a_stacked,
|
|
lora_b_stacked,
|
|
topk_weights,
|
|
_sorted,
|
|
_eids,
|
|
num_tokens_post_padded_lora,
|
|
max_lora_rank,
|
|
top_k,
|
|
shrink_config,
|
|
expand_config,
|
|
adapter_enabled,
|
|
fully_sharded=fully_sharded,
|
|
token_lora_mapping=token_lora_mapping,
|
|
add_inputs=add_inputs,
|
|
)
|
|
|
|
return (
|
|
sorted_token_ids_lora,
|
|
expert_ids_lora,
|
|
num_tokens_post_padded_lora,
|
|
token_lora_mapping,
|
|
)
|
|
|
|
def add_lora_w2(
|
|
self,
|
|
y: torch.Tensor,
|
|
x: torch.Tensor,
|
|
lora_a_stacked: tuple[torch.Tensor, ...],
|
|
lora_b_stacked: tuple[torch.Tensor, ...],
|
|
topk_weights: torch.Tensor,
|
|
sorted_token_ids_lora: torch.Tensor | None,
|
|
expert_ids_lora: torch.Tensor | None,
|
|
num_tokens_post_padded_lora: torch.Tensor | None,
|
|
token_lora_mapping: torch.Tensor | None,
|
|
num_tokens: int,
|
|
w1: torch.Tensor,
|
|
w2: torch.Tensor,
|
|
top_k_num: int,
|
|
max_loras: int,
|
|
adapter_enabled: torch.Tensor,
|
|
top_k: int,
|
|
fully_sharded: bool,
|
|
tp_rank: int,
|
|
use_tuned_config: bool,
|
|
add_inputs: bool = True,
|
|
) -> None:
|
|
import functools
|
|
|
|
from vllm.lora.layers.utils import try_get_optimal_moe_lora_config
|
|
from vllm.lora.ops.triton_ops.utils import (
|
|
_normalize_lora_config_keys,
|
|
get_lora_op_configs,
|
|
)
|
|
from vllm.model_executor.layers.fused_moe.config import _get_config_dtype_str
|
|
|
|
config_dtype = _get_config_dtype_str(
|
|
dtype=x.dtype,
|
|
use_fp8_w8a8=False,
|
|
use_int8_w8a16=False,
|
|
use_int4_w4a16=False,
|
|
)
|
|
max_lora_rank = lora_a_stacked[0].shape[-2]
|
|
|
|
if use_tuned_config:
|
|
shrink_config = get_lora_op_configs(
|
|
op_type="fused_moe_lora_w2_shrink",
|
|
max_loras=max_loras,
|
|
batch=num_tokens,
|
|
hidden_size=y.shape[-1],
|
|
rank=max_lora_rank,
|
|
num_slices=1,
|
|
moe_intermediate_size=lora_a_stacked[0].shape[-1],
|
|
)
|
|
expand_config = get_lora_op_configs(
|
|
op_type="fused_moe_lora_w2_expand",
|
|
max_loras=max_loras,
|
|
batch=num_tokens,
|
|
hidden_size=y.shape[-1],
|
|
rank=max_lora_rank,
|
|
num_slices=1,
|
|
moe_intermediate_size=lora_a_stacked[0].shape[-1],
|
|
)
|
|
else:
|
|
get_config = functools.partial(
|
|
try_get_optimal_moe_lora_config,
|
|
w1_shape=w1.shape,
|
|
w2_shape=w2.shape,
|
|
rank=max_lora_rank,
|
|
top_k=top_k,
|
|
dtype=config_dtype,
|
|
M=num_tokens,
|
|
)
|
|
shrink_config = get_config(op_type="fused_moe_lora_w2_shrink")
|
|
expand_config = get_config(op_type="fused_moe_lora_w2_expand")
|
|
|
|
shrink_config = _normalize_lora_config_keys(shrink_config)
|
|
expand_config = _normalize_lora_config_keys(expand_config)
|
|
|
|
_sorted = sorted_token_ids_lora
|
|
_eids = expert_ids_lora
|
|
if _sorted is not None:
|
|
assert _eids is not None
|
|
_eids = _eids.view(max_loras, -1)
|
|
_sorted = _sorted.view(max_loras, -1)
|
|
|
|
# w2_lora_b shape[-2] is hidden_size // tp_size when fully_sharded
|
|
shard_size = lora_b_stacked[0].shape[-2]
|
|
offset = shard_size * tp_rank if fully_sharded else 0
|
|
|
|
self.add_lora_fused_moe(
|
|
y,
|
|
x,
|
|
lora_a_stacked,
|
|
lora_b_stacked,
|
|
topk_weights,
|
|
_sorted,
|
|
_eids,
|
|
num_tokens_post_padded_lora,
|
|
max_lora_rank,
|
|
top_k,
|
|
shrink_config,
|
|
expand_config,
|
|
adapter_enabled,
|
|
True, # mul_routed_weight
|
|
fully_sharded=fully_sharded,
|
|
offset=offset,
|
|
token_lora_mapping=token_lora_mapping,
|
|
add_inputs=add_inputs,
|
|
)
|