Skip to content

GitLab

Unverified Commit e1060448 authored by Kunshang Ji's avatar Kunshang Ji Committed by GitHub
Browse files

[XPU][1/N] Deprecate ipex and switch to vllm-xpu-kernels for xpu platform (#33379) 


Signed-off-by: default avatarKunshang Ji <kunshang.ji@intel.com>
parent bf001da4
Show whitespace changes
Inline Side-by-side
Showing with 144 additions and 921 deletions
.buildkite/scripts/hardware_ci/run-xpu-test.sh
View file @ e1060448
...@@ -38,10 +38,9 @@ docker run \ ...@@ -38,10 +38,9 @@ docker run \
python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
python3 examples/offline_inference/basic/generate.py --model Intel/Qwen2.5-0.5B-W4A16-G128-AutoRound-LLMC-TEST-ONLY --enforce-eager
python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
cd tests cd tests
pytest -v -s v1/core pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py
pytest -v -s v1/engine pytest -v -s v1/engine
pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
......
docker/Dockerfile.xpu
View file @ e1060448
FROM intel/deep-learning-essentials:2025.2.2-0-devel-ubuntu24.04 AS vllm-base FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04 AS vllm-base
RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \ echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
add-apt-repository -y ppa:kobuk-team/intel-graphics-staging add-apt-repository -y ppa:kobuk-team/intel-graphics
RUN apt clean && apt-get update -y && \ RUN apt clean && apt-get update -y && \
apt-get install -y --no-install-recommends --fix-missing \ apt-get install -y --no-install-recommends --fix-missing \
...@@ -25,10 +25,13 @@ RUN apt clean && apt-get update -y && \ ...@@ -25,10 +25,13 @@ RUN apt clean && apt-get update -y && \
RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 1 RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 1
RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.12 1 RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.12 1
RUN apt install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd libze-intel-gpu-raytracing intel-ocloc RUN apt update && apt upgrade -y && \
apt install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd libze-intel-gpu-raytracing intel-ocloc && \
apt install -y intel-oneapi-compiler-dpcpp-cpp-2025.3
# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2. # This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2.
ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.6_offline.sh" ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.8_offline.sh"
RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.7/${ONECCL_INSTALLER}" && \ RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.7/${ONECCL_INSTALLER}" && \
bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \ bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
rm "${ONECCL_INSTALLER}" && \ rm "${ONECCL_INSTALLER}" && \
...@@ -85,6 +88,9 @@ RUN python3 -m pip install -e tests/vllm_test_utils ...@@ -85,6 +88,9 @@ RUN python3 -m pip install -e tests/vllm_test_utils
ENV NIXL_VERSION=0.7.0 ENV NIXL_VERSION=0.7.0
RUN python3 /workspace/vllm/tools/install_nixl_from_source_ubuntu.py RUN python3 /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
# FIX triton
RUN --mount=type=cache,target=/root/.cache/pip pip uninstall triton triton-xpu -y && pip install triton-xpu==3.6.0 --extra-index-url=https://download.pytorch.org/whl/xpu
# PyJWT-2.7.0 will influence some wheel behaviors, remove its dist-info to avoid conflicts # PyJWT-2.7.0 will influence some wheel behaviors, remove its dist-info to avoid conflicts
RUN rm /usr/lib/python3/dist-packages/PyJWT-2.7.0.dist-info/ -rf RUN rm /usr/lib/python3/dist-packages/PyJWT-2.7.0.dist-info/ -rf
......
requirements/xpu.txt
View file @ e1060448
...@@ -11,8 +11,8 @@ jinja2>=3.1.6 ...@@ -11,8 +11,8 @@ jinja2>=3.1.6
datasets # for benchmark scripts datasets # for benchmark scripts
numba == 0.61.2 # Required for N-gram speculative decoding numba == 0.61.2 # Required for N-gram speculative decoding
--extra-index-url=https://download.pytorch.org/whl/xpu --extra-index-url=https://download.pytorch.org/whl/xpu
torch==2.9.0+xpu torch==2.10.0+xpu
torchaudio torchaudio
torchvision torchvision
intel-extension-for-pytorch @ https://intel-extension-for-pytorch.s3.us-east-1.amazonaws.com/ipex_dev/xpu/intel_extension_for_pytorch-2.9.10.post0%2Bxpu-cp312-cp312-linux_x86_64.whl vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.0/vllm_xpu_kernels-0.1.0-cp312-cp312-linux_x86_64.whl
\ No newline at end of file
vllm/_ipex_ops.py
View file @ e1060448
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import TYPE_CHECKING
import torch import torch
from vllm_xpu_kernels.flash_attn_interface import flash_attn_varlen_func
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.platforms import current_platform
logger = init_logger(__name__) logger = init_logger(__name__)
try: if TYPE_CHECKING:
import intel_extension_for_pytorch as ipex
except ImportError as e:
logger.debug("Import error msg: %s", e.msg)
def register_fake(fn):
return lambda name: fn
else:
try:
from torch.library import register_fake
except ImportError:
from torch.library import impl_abstract as register_fake
class ipex_ops: if hasattr(torch.ops._xpu_C, "fp8_gemm_w8a16"):
@staticmethod
def _reshape_activation_tensor(
x: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
num = x.size(0)
d = x.size(1) // 2
x = x.reshape(num, 2, d)
x1, x2 = torch.chunk(x, chunks=2, dim=1)
x1 = x1.reshape(num, d)
x2 = x2.reshape(num, d)
return x1, x2
@staticmethod
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
ipex.llm.functional.silu_and_mul(x, out)
@staticmethod
def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
ipex.llm.functional.gelu_and_mul(x, out)
@staticmethod
def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
ipex.llm.functional.gelu_and_mul(x, out)
@staticmethod
def gelu_fast(x: torch.Tensor) -> torch.Tensor:
return torch.nn.functional.gelu(x)
@staticmethod @register_fake("_xpu_C::fp8_gemm_w8a16")
def gelu_new(x: torch.Tensor) -> torch.Tensor: def _fp8_gemm_w8a16_fake(
return torch.nn.functional.gelu(x) input: torch.Tensor,
q_weight: torch.Tensor,
@staticmethod weight_scale: torch.Tensor,
def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None: bias: torch.Tensor | None = None,
ipex.llm.functional.gelu_quick(x, out)
@staticmethod
def paged_attention_v1(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
context_lens: torch.Tensor,
block_size: int,
max_context_len: int,
alibi_slopes: torch.Tensor | None,
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0,
) -> None:
assert kv_cache_dtype == "auto"
num_heads = out.size(1)
num_queries_per_tokens = num_heads // num_kv_heads
ipex.llm.modules.PagedAttention.single_query_kv_attention(
out,
query.contiguous(),
key_cache.view_as(value_cache),
value_cache,
num_queries_per_tokens,
scale,
block_tables,
context_lens,
block_size,
max_context_len,
alibi_slopes,
)
@staticmethod
def paged_attention_v2(
out: torch.Tensor,
exp_sum: torch.Tensor,
max_logits: torch.Tensor,
tmp_out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
num_kv_heads: int,
scale: float,
block_tables: torch.Tensor,
context_lens: torch.Tensor,
block_size: int,
max_context_len: int,
alibi_slopes: torch.Tensor | None,
kv_cache_dtype: str,
k_scale: float,
v_scale: float,
tp_rank: int = 0,
blocksparse_local_blocks: int = 0,
blocksparse_vert_stride: int = 0,
blocksparse_block_size: int = 64,
blocksparse_head_sliding_step: int = 0,
) -> None:
assert kv_cache_dtype == "auto"
num_heads = out.size(1)
num_queries_per_tokens = num_heads // num_kv_heads
ipex.llm.modules.PagedAttention.single_query_kv_attention(
out,
query.contiguous(),
key_cache.view_as(value_cache),
value_cache,
num_queries_per_tokens,
scale,
block_tables,
context_lens,
block_size,
max_context_len,
alibi_slopes,
)
@staticmethod
def rotary_embedding(
positions: torch.Tensor, # [batch_size, seq_len]
query: torch.Tensor, # [batch_size, seq_len, num_heads*head_size]
key: torch.Tensor, # [batch_size, seq_len, num_kv_heads*head_size]
head_size: int,
cos_sin_cache: torch.Tensor, # [cos_sin_dim, rot_dim]
is_neox: bool,
) -> None:
rot_dim = cos_sin_cache.size(1)
ipex.llm.functional.rotary_embedding_batched(
positions, query, key, head_size, cos_sin_cache, is_neox, rot_dim
)
@staticmethod
def rms_norm(
input: torch.Tensor, weight: torch.Tensor, epsilon: float
) -> torch.Tensor: ) -> torch.Tensor:
out = torch.empty_like(input) input_2d = input.view(-1, input.shape[-1])
torch.ops.torch_ipex.rms_norm_vllm(out, input.contiguous(), weight, epsilon) M = input_2d.size(0)
return out N = q_weight.size(1)
return torch.empty((M, N), dtype=input.dtype, device=input.device)
@staticmethod
def fused_add_rms_norm(
input: torch.Tensor,
residual: torch.Tensor,
weight: torch.Tensor,
epsilon: float,
) -> None:
torch.ops.torch_ipex.fused_add_rms_norm_vllm(input, residual, weight, epsilon)
@staticmethod if hasattr(torch.ops._xpu_C, "int4_gemm_w4a16"):
def varlen_attention(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
out: torch.Tensor,
seqlen_q: torch.Tensor,
seqlen_k: torch.Tensor,
alibi_slopes: torch.Tensor | None,
max_seqlen_q: int,
max_seqlen_k: int,
pdropout: float,
softmax_scale: float,
zero_tensors: bool,
is_causal: bool,
return_softmax: bool,
gen_: torch.Generator,
window_size_left: float,
window_size_right: float,
logits_soft_cap: float,
) -> None:
if ipex.__version__.endswith("cpu"):
if logits_soft_cap != 0.0:
raise ValueError("IPEX CPU does not support logits_soft_cap")
assert alibi_slopes is None
assert window_size_left < 0 and window_size_right < 0
ipex.llm.functional.varlen_attention(
query.contiguous(),
key.contiguous(),
value.contiguous(),
out,
seqlen_q.int(),
seqlen_k.int(),
max_seqlen_q,
max_seqlen_k,
pdropout,
softmax_scale,
zero_tensors,
is_causal,
return_softmax,
gen_,
)
else: # XPU build
ipex.llm.functional.varlen_attention(
query.contiguous(),
key.contiguous(),
value.contiguous(),
out,
seqlen_q.int(),
seqlen_k.int(),
alibi_slopes,
max_seqlen_q,
max_seqlen_k,
pdropout,
softmax_scale,
zero_tensors,
is_causal,
return_softmax,
gen_,
window_size_left,
window_size_right,
logits_soft_cap,
)
@staticmethod @register_fake("_xpu_C::int4_gemm_w4a16")
def reshape_and_cache( def _int4_gemm_w4a16_fake(
key: torch.Tensor, input: torch.Tensor,
value: torch.Tensor, q_weight: torch.Tensor,
key_cache: torch.Tensor, bias: torch.Tensor | None,
value_cache: torch.Tensor, weight_scale: torch.Tensor,
slot_mapping: torch.Tensor, qzeros: torch.Tensor,
kv_cache_dtype: str, group_size: int,
k_scale: float, group_idx: torch.Tensor | None = None,
v_scale: float, ) -> torch.Tensor:
) -> None: input_2d = input.view(-1, input.shape[-1])
assert kv_cache_dtype == "auto" M = input_2d.size(0)
ipex.llm.modules.PagedAttention.reshape_and_cache( N = q_weight.size(1)
key, value, key_cache, value_cache, slot_mapping return torch.empty((M, N), dtype=input.dtype, device=input.device)
)
@staticmethod
def reshape_and_cache_flash(
key: torch.Tensor,
value: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
slot_mapping: torch.Tensor,
kv_cache_dtype: str,
k_scale: torch.Tensor | None = None,
v_scale: torch.Tensor | None = None,
k_scale_float: float = 1.0,
v_scale_float: float = 1.0,
) -> None:
ipex.llm.modules.PagedAttention.reshape_and_cache_flash(
key,
value,
key_cache,
value_cache,
slot_mapping,
kv_cache_dtype,
k_scale_float,
v_scale_float,
)
class ipex_ops:
@staticmethod @staticmethod
def flash_attn_varlen_func( def flash_attn_varlen_func(
q: torch.Tensor, q: torch.Tensor,
...@@ -295,8 +81,21 @@ class ipex_ops: ...@@ -295,8 +81,21 @@ class ipex_ops:
k_descale=None, k_descale=None,
v_descale=None, v_descale=None,
num_splits=0, num_splits=0,
return_softmax_lse: bool | None = False,
s_aux: torch.Tensor | None = None, s_aux: torch.Tensor | None = None,
): ):
assert cu_seqlens_k is not None or seqused_k is not None, (
"cu_seqlens_k or seqused_k must be provided"
)
assert cu_seqlens_k is None or seqused_k is None, (
"cu_seqlens_k and seqused_k cannot be provided at the same time"
)
assert block_table is None or seqused_k is not None, (
"when enable block_table, seqused_k is needed"
)
assert block_table is not None or cu_seqlens_k is not None, (
"when block_table is disabled, cu_seqlens_k is needed"
)
if out is None: if out is None:
out = torch.empty(q.shape, dtype=q.dtype, device=q.device) out = torch.empty(q.shape, dtype=q.dtype, device=q.device)
real_window_size: tuple[int, int] real_window_size: tuple[int, int]
...@@ -304,55 +103,30 @@ class ipex_ops: ...@@ -304,55 +103,30 @@ class ipex_ops:
real_window_size = (-1, -1) real_window_size = (-1, -1)
else: else:
assert len(window_size) == 2 assert len(window_size) == 2
real_window_size = (window_size[0], window_size[1]) real_window_size = (window_size[0], window_size[1]) # noqa: F841
# In encode attention, v maybe not contiguous and current
# kernel can't handle it
if block_table is None: if block_table is None:
assert cu_seqlens_k is not None, ( v = v.contiguous()
"cu_seqlens_k can't be None when calling varlen_attention." return flash_attn_varlen_func(
) out=out,
if softmax_scale is None: q=q.contiguous(),
softmax_scale = q.shape[-1] ** (-0.5) k=k,
ipex_ops.varlen_attention( v=v,
q.contiguous(), cu_seqlens_q=cu_seqlens_q,
k.contiguous(), cu_seqlens_k=cu_seqlens_k,
v.contiguous(), seqused_k=seqused_k,
out, max_seqlen_q=max_seqlen_q,
cu_seqlens_q, max_seqlen_k=max_seqlen_k,
cu_seqlens_k, softmax_scale=softmax_scale,
None, causal=causal,
max_seqlen_q, block_table=block_table,
max_seqlen_k, s_aux=s_aux,
0.0, window_size=real_window_size,
softmax_scale, # alibi_slopes = alibi_slopes,
False, # softcap=softcap,
causal, return_softmax_lse=return_softmax_lse,
False,
None,
real_window_size[0],
real_window_size[1],
-1,
)
return out
else:
return ipex.llm.modules.PagedAttention.flash_attn_varlen_func(
out,
q.contiguous(),
k,
v,
cu_seqlens_q,
seqused_k,
max_seqlen_q,
max_seqlen_k,
softmax_scale,
causal,
block_table,
alibi_slopes,
sink=s_aux,
softcap=softcap,
window_size_left=real_window_size[0],
window_size_right=real_window_size[1],
k_scale=1.0,
v_scale=1.0,
) )
@staticmethod @staticmethod
...@@ -382,64 +156,3 @@ class ipex_ops: ...@@ -382,64 +156,3 @@ class ipex_ops:
"get_scheduler_metadata is not implemented for ipex_ops, returning None." "get_scheduler_metadata is not implemented for ipex_ops, returning None."
) )
return None return None
@staticmethod
def swap_blocks(
src: torch.Tensor, dst: torch.Tensor, block_mapping: torch.Tensor
) -> None:
torch.xpu.swap_blocks(src, dst, block_mapping) # type: ignore
@staticmethod
def scaled_fp8_quant(
input: torch.Tensor,
scale: torch.Tensor | None = None,
num_token_padding: int | None = None,
scale_ub: torch.Tensor | None = None,
use_per_token_if_dynamic: bool = False,
output: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Quantize input tensor to FP8 and return quantized tensor and scale.
This function is designed for both static and dynamic quantization:
If you provide the scale, it will use static scaling and if you omit
it, the scale will be determined dynamically. Currently, XPU platform
only supports dynamic quantization. The function also allows optional
padding of the output tensors for downstream kernels that will benefit
from padding.
Args:
input: The input tensor to be quantized to FP8
scale: Optional scaling factor for the FP8 quantization
scale_ub: Optional upper bound for scaling factor in dynamic
per token case
num_token_padding: If specified, pad the first dimension
of the output to at least this value.
use_per_token_if_dynamic: Whether to do per_tensor or per_token
in the dynamic quantization case.
Returns:
tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
scaling factor.
"""
# This code assumes batch_dim and num_tokens are flattened
assert input.ndim == 2
shape: tuple[int, int] | torch.Size = input.shape
out_dtype: torch.dtype = current_platform.fp8_dtype()
if num_token_padding:
shape = (max(num_token_padding, input.shape[0]), shape[1])
if output is None:
output = torch.empty(shape, device=input.device, dtype=out_dtype)
else:
assert num_token_padding is None, (
"padding not supported if output passed in"
)
assert output.dtype == out_dtype
assert scale is None, "only dynamic fp8 quantization supported on XPU"
assert not use_per_token_if_dynamic, (
"per token dynamic fp8 quantization not supported on XPU"
)
scale = torch.zeros(1, device=input.device, dtype=torch.float32)
torch.ops.torch_ipex.dynamic_scaled_fp8_quant(output, input, scale)
return output, scale
vllm/config/model.py
View file @ e1060448
...@@ -877,7 +877,6 @@ class ModelConfig: ...@@ -877,7 +877,6 @@ class ModelConfig:
overrides = [ overrides = [
"gptq_marlin", "gptq_marlin",
"awq_marlin", "awq_marlin",
"ipex",
"inc", "inc",
"moe_wna16", "moe_wna16",
"modelopt", "modelopt",
......
vllm/model_executor/layers/activation.py
View file @ e1060448
...@@ -129,12 +129,8 @@ class SiluAndMul(CustomOp): ...@@ -129,12 +129,8 @@ class SiluAndMul(CustomOp):
def __init__(self, *, compile_native: bool = True): def __init__(self, *, compile_native: bool = True):
super().__init__(compile_native=compile_native) super().__init__(compile_native=compile_native)
if current_platform.is_cuda_alike(): if current_platform.is_cuda_alike() or current_platform.is_xpu():
self.op = torch.ops._C.silu_and_mul self.op = torch.ops._C.silu_and_mul
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
self.op = ipex_ops.silu_and_mul
elif current_platform.is_cpu(): elif current_platform.is_cpu():
self._forward_method = self.forward_native self._forward_method = self.forward_native
...@@ -152,11 +148,7 @@ class SiluAndMul(CustomOp): ...@@ -152,11 +148,7 @@ class SiluAndMul(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2 return self.forward_cuda(x)
output_shape = x.shape[:-1] + (d,)
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
self.op(out, x)
return out
# --8<-- [start:mul_and_silu] # --8<-- [start:mul_and_silu]
...@@ -175,12 +167,8 @@ class MulAndSilu(CustomOp): ...@@ -175,12 +167,8 @@ class MulAndSilu(CustomOp):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
if current_platform.is_cuda_alike(): if current_platform.is_cuda_alike() or current_platform.is_xpu():
self.op = torch.ops._C.mul_and_silu self.op = torch.ops._C.mul_and_silu
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
self.op = ipex_ops.silu_and_mul
elif current_platform.is_cpu(): elif current_platform.is_cpu():
self._forward_method = self.forward_native self._forward_method = self.forward_native
...@@ -196,8 +184,8 @@ class MulAndSilu(CustomOp): ...@@ -196,8 +184,8 @@ class MulAndSilu(CustomOp):
self.op(out, x) self.op(out, x)
return out return out
# TODO implement forward_xpu for MulAndSilu def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
# def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: return self.forward_cuda(x)
# --8<-- [start:gelu_and_mul_sparse] # --8<-- [start:gelu_and_mul_sparse]
...@@ -278,7 +266,11 @@ class GeluAndMul(CustomOp): ...@@ -278,7 +266,11 @@ class GeluAndMul(CustomOp):
self.approximate = approximate self.approximate = approximate
if approximate not in ("none", "tanh"): if approximate not in ("none", "tanh"):
raise ValueError(f"Unknown approximate mode: {approximate}") raise ValueError(f"Unknown approximate mode: {approximate}")
if current_platform.is_cuda_alike() or current_platform.is_cpu(): if (
current_platform.is_cuda_alike()
or current_platform.is_cpu()
or current_platform.is_xpu()
):
if approximate == "none": if approximate == "none":
self.op = torch.ops._C.gelu_and_mul self.op = torch.ops._C.gelu_and_mul
elif approximate == "tanh": elif approximate == "tanh":
...@@ -289,13 +281,6 @@ class GeluAndMul(CustomOp): ...@@ -289,13 +281,6 @@ class GeluAndMul(CustomOp):
"with torch.compile. For native implementation, fallback to 'none' " "with torch.compile. For native implementation, fallback to 'none' "
"approximation. The custom kernel implementation is unaffected." "approximation. The custom kernel implementation is unaffected."
) )
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
if approximate == "none":
self.op = ipex_ops.gelu_and_mul
else:
self.op = ipex_ops.gelu_tanh_and_mul
def forward_native(self, x: torch.Tensor) -> torch.Tensor: def forward_native(self, x: torch.Tensor) -> torch.Tensor:
"""PyTorch-native implementation equivalent to forward().""" """PyTorch-native implementation equivalent to forward()."""
...@@ -314,11 +299,7 @@ class GeluAndMul(CustomOp): ...@@ -314,11 +299,7 @@ class GeluAndMul(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2 return self.forward_cuda(x)
output_shape = x.shape[:-1] + (d,)
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
self.op(out, x)
return out
def extra_repr(self) -> str: def extra_repr(self) -> str:
return f"approximate={repr(self.approximate)}" return f"approximate={repr(self.approximate)}"
...@@ -401,12 +382,12 @@ class NewGELU(CustomOp): ...@@ -401,12 +382,12 @@ class NewGELU(CustomOp):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
if current_platform.is_cuda_alike() or current_platform.is_cpu(): if (
current_platform.is_cuda_alike()
or current_platform.is_cpu()
or current_platform.is_xpu()
):
self.op = torch.ops._C.gelu_new self.op = torch.ops._C.gelu_new
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
self.op = ipex_ops.gelu_new
def forward_native(self, x: torch.Tensor) -> torch.Tensor: def forward_native(self, x: torch.Tensor) -> torch.Tensor:
"""PyTorch-native implementation equivalent to forward().""" """PyTorch-native implementation equivalent to forward()."""
...@@ -419,7 +400,7 @@ class NewGELU(CustomOp): ...@@ -419,7 +400,7 @@ class NewGELU(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
return self.op(x) return self.forward_cuda(x)
# --8<-- [start:gelu_fast] # --8<-- [start:gelu_fast]
...@@ -429,12 +410,12 @@ class FastGELU(CustomOp): ...@@ -429,12 +410,12 @@ class FastGELU(CustomOp):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
if current_platform.is_cuda_alike() or current_platform.is_cpu(): if (
current_platform.is_cuda_alike()
or current_platform.is_cpu()
or current_platform.is_xpu()
):
self.op = torch.ops._C.gelu_fast self.op = torch.ops._C.gelu_fast
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
self.op = ipex_ops.gelu_fast
def forward_native(self, x: torch.Tensor) -> torch.Tensor: def forward_native(self, x: torch.Tensor) -> torch.Tensor:
"""PyTorch-native implementation equivalent to forward().""" """PyTorch-native implementation equivalent to forward()."""
...@@ -446,7 +427,7 @@ class FastGELU(CustomOp): ...@@ -446,7 +427,7 @@ class FastGELU(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
return self.op(x) return self.forward_cuda(x)
# --8<-- [start:quick_gelu] # --8<-- [start:quick_gelu]
...@@ -457,12 +438,12 @@ class QuickGELU(CustomOp): ...@@ -457,12 +438,12 @@ class QuickGELU(CustomOp):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
if current_platform.is_cuda_alike() or current_platform.is_cpu(): if (
current_platform.is_cuda_alike()
or current_platform.is_cpu()
or current_platform.is_xpu()
):
self.op = torch.ops._C.gelu_quick self.op = torch.ops._C.gelu_quick
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops
self.op = ipex_ops.gelu_quick
def forward_native(self, x: torch.Tensor) -> torch.Tensor: def forward_native(self, x: torch.Tensor) -> torch.Tensor:
"""PyTorch-native implementation equivalent to forward().""" """PyTorch-native implementation equivalent to forward()."""
...@@ -474,12 +455,7 @@ class QuickGELU(CustomOp): ...@@ -474,12 +455,7 @@ class QuickGELU(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
out = torch.empty_like(x) return self.forward_cuda(x)
self.op(out, x)
return out
# TODO implement forward_xpu for QuickGELU
# def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
# --8<-- [start:relu2] # --8<-- [start:relu2]
......
vllm/model_executor/layers/layernorm.py
View file @ e1060448
...@@ -231,24 +231,7 @@ class RMSNorm(CustomOp): ...@@ -231,24 +231,7 @@ class RMSNorm(CustomOp):
x: torch.Tensor, x: torch.Tensor,
residual: torch.Tensor | None = None, residual: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if self.variance_size_override is not None: return self.forward_cuda(x, residual)
return self.forward_native(x, residual)
from vllm._ipex_ops import ipex_ops as ops
if residual is not None:
ops.fused_add_rms_norm(
x,
residual,
self.weight.data,
self.variance_epsilon,
)
return x, residual
return ops.rms_norm(
x,
self.weight.data,
self.variance_epsilon,
)
def extra_repr(self) -> str: def extra_repr(self) -> str:
s = f"hidden_size={self.weight.data.size(0)}" s = f"hidden_size={self.weight.data.size(0)}"
......
vllm/model_executor/layers/linear.py
View file @ e1060448
...@@ -60,8 +60,6 @@ WEIGHT_LOADER_V2_SUPPORTED = [ ...@@ -60,8 +60,6 @@ WEIGHT_LOADER_V2_SUPPORTED = [
"ModelOptFp8LinearMethod", "ModelOptFp8LinearMethod",
"ModelOptFp8PcPtLinearMethod", "ModelOptFp8PcPtLinearMethod",
"ModelOptFp8PbWoLinearMethod", "ModelOptFp8PbWoLinearMethod",
"IPEXAWQLinearMethod",
"IPEXGPTQLinearMethod",
"QuarkLinearMethod", "QuarkLinearMethod",
"ModelOptNvFp4LinearMethod", "ModelOptNvFp4LinearMethod",
"PetitNvFp4LinearMethod", "PetitNvFp4LinearMethod",
......
vllm/model_executor/layers/quantization/__init__.py
View file @ e1060448
...@@ -24,7 +24,6 @@ QuantizationMethods = Literal[ ...@@ -24,7 +24,6 @@ QuantizationMethods = Literal[
"compressed-tensors", "compressed-tensors",
"bitsandbytes", "bitsandbytes",
"experts_int8", "experts_int8",
"ipex",
"quark", "quark",
"moe_wna16", "moe_wna16",
"torchao", "torchao",
...@@ -41,7 +40,6 @@ DEPRECATED_QUANTIZATION_METHODS = [ ...@@ -41,7 +40,6 @@ DEPRECATED_QUANTIZATION_METHODS = [
"fbgemm_fp8", "fbgemm_fp8",
"fp_quant", "fp_quant",
"experts_int8", "experts_int8",
"ipex",
"petit_nvfp4", "petit_nvfp4",
] ]
...@@ -121,7 +119,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]: ...@@ -121,7 +119,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
from .gptq import GPTQConfig from .gptq import GPTQConfig
from .gptq_marlin import GPTQMarlinConfig from .gptq_marlin import GPTQMarlinConfig
from .inc import INCConfig from .inc import INCConfig
from .ipex_quant import IPEXConfig
from .modelopt import ModelOptFp8Config, ModelOptNvFp4Config from .modelopt import ModelOptFp8Config, ModelOptNvFp4Config
from .moe_wna16 import MoeWNA16Config from .moe_wna16 import MoeWNA16Config
from .mxfp4 import Mxfp4Config from .mxfp4 import Mxfp4Config
...@@ -144,7 +141,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]: ...@@ -144,7 +141,6 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
"bitsandbytes": BitsAndBytesConfig, "bitsandbytes": BitsAndBytesConfig,
"ptpc_fp8": PTPCFp8Config, "ptpc_fp8": PTPCFp8Config,
"experts_int8": ExpertsInt8Config, "experts_int8": ExpertsInt8Config,
"ipex": IPEXConfig,
"quark": QuarkConfig, "quark": QuarkConfig,
"moe_wna16": MoeWNA16Config, "moe_wna16": MoeWNA16Config,
"torchao": TorchAOConfig, "torchao": TorchAOConfig,
......
vllm/model_executor/layers/quantization/fp8.py
View file @ e1060448
...@@ -184,39 +184,10 @@ class Fp8Config(QuantizationConfig): ...@@ -184,39 +184,10 @@ class Fp8Config(QuantizationConfig):
def get_xpu_quant_method( def get_xpu_quant_method(
self, layer: torch.nn.Module, prefix: str self, layer: torch.nn.Module, prefix: str
) -> "QuantizeMethodBase | None": ) -> "QuantizeMethodBase | None":
from vllm.model_executor.layers.quantization.ipex_quant import ( raise NotImplementedError(
XPUFp8LinearMethod, "FP8 quantization is not supported during xpu kernel migration."
XPUFp8MoEMethod,
) )
fp8_config = Fp8Config(
is_checkpoint_fp8_serialized=self.is_checkpoint_fp8_serialized,
activation_scheme=self.activation_scheme,
ignored_layers=self.ignored_layers,
weight_block_size=self.weight_block_size,
)
if isinstance(layer, LinearBase):
if is_layer_skipped(
prefix=prefix,
ignored_layers=self.ignored_layers,
fused_mapping=self.packed_modules_mapping,
):
return UnquantizedLinearMethod()
return XPUFp8LinearMethod(fp8_config)
elif isinstance(layer, FusedMoE):
if is_layer_skipped(
prefix=prefix,
ignored_layers=self.ignored_layers,
fused_mapping=self.packed_modules_mapping,
):
return UnquantizedFusedMoEMethod(layer.moe_config)
return XPUFp8MoEMethod(fp8_config, layer)
elif isinstance(layer, Attention):
return Fp8KVCacheMethod(self)
return None
def get_quant_method( def get_quant_method(
self, layer: torch.nn.Module, prefix: str self, layer: torch.nn.Module, prefix: str
) -> "QuantizeMethodBase | None": ) -> "QuantizeMethodBase | None":
......
vllm/model_executor/layers/quantization/inc.py
View file @ e1060448
...@@ -38,7 +38,6 @@ class INCConfig(QuantizationConfig): ...@@ -38,7 +38,6 @@ class INCConfig(QuantizationConfig):
"awq", "awq",
"awq:marlin", "awq:marlin",
"marlin", "marlin",
"ipex",
} }
def __init__( def __init__(
...@@ -410,31 +409,10 @@ class INCConfig(QuantizationConfig): ...@@ -410,31 +409,10 @@ class INCConfig(QuantizationConfig):
return UnquantizedLinearMethod() return UnquantizedLinearMethod()
else: else:
return None return None
from vllm.model_executor.layers.quantization.ipex_quant import ( raise NotImplementedError(
IPEXAWQLinearMethod, "INC quantization is not supported during xpu kernel migration."
IPEXConfig,
IPEXGPTQLinearMethod,
) )
if isinstance(layer, (LinearBase, ParallelLMHead)):
if "awq" in self.packing_format:
config = IPEXConfig(
method="awq", weight_bits=weight_bits, group_size=group_size
)
return IPEXAWQLinearMethod(config)
elif "gptq" in self.packing_format:
config = IPEXConfig(
method="gptq", weight_bits=weight_bits, group_size=group_size
)
return IPEXGPTQLinearMethod(config)
else:
raise ValueError(
f"ipex backend only supports awq "
f"and gptq format,but got {self.packing_format}"
)
else:
return None
def get_quant_method(self, layer: torch.nn.Module, prefix: str): def get_quant_method(self, layer: torch.nn.Module, prefix: str):
if prefix and self.extra_config: if prefix and self.extra_config:
for layer_name in self.extra_config: for layer_name in self.extra_config:
......
vllm/model_executor/layers/quantization/ipex_quant.py deleted 100644 → 0
View file @ bf001da4
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
import torch
from packaging import version
from torch.nn import Module
from vllm._ipex_ops import ipex_ops as ops
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.linear import (
LinearBase,
LinearMethodBase,
UnquantizedLinearMethod,
)
from vllm.model_executor.layers.quantization import (
QuantizationConfig,
QuantizationMethods,
)
from vllm.model_executor.layers.quantization.awq import AWQLinearMethod
from vllm.model_executor.layers.quantization.fp8 import (
Fp8Config,
Fp8LinearMethod,
Fp8OnlineMoEMethod,
)
from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
MIN_IPEX_VERSION = "2.6.0"
class IPEXConfig(QuantizationConfig):
"""INT8 quantization config class using IPEX for the CPU/XPU backend,
including AWQ, GPTQ.
"""
IPEX_QUANT_METHOD_MAP = {
"awq": 1,
"gptq": 0,
}
def __init__(
self,
method: str,
weight_bits: int,
group_size: int,
modules_to_not_convert: list[str] | None = None,
desc_act: bool | None = None,
lm_head_quantized: bool | None = None,
is_sym: bool | None = None,
) -> None:
super().__init__()
self.method = method
self.weight_bits = weight_bits
self.group_size = group_size
self.modules_to_not_convert = modules_to_not_convert or []
self.desc_act = desc_act
self.lm_head_quantized = lm_head_quantized
self.is_sym = is_sym
self.pack_factor = 32 // self.weight_bits
if self.weight_bits not in [4]:
raise ValueError(
f"IPEX quantization supports weight bits [4], "
f"but got {self.weight_bits}."
)
if self.method not in ["awq", "gptq"]:
raise ValueError(
f"IPEX quantization supports [awq, gptq], but got {self.method}."
)
def __repr__(self) -> str:
return (
f"IPEXConfig(method={self.method},"
f"weight_bits={self.weight_bits}, "
f"group_size={self.group_size})"
)
@classmethod
def get_name(cls) -> QuantizationMethods:
return "ipex"
@classmethod
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
return [torch.bfloat16, torch.float16]
@classmethod
def get_min_capability(cls) -> int:
return -1
@staticmethod
def get_config_filenames() -> list[str]:
return [
"quant_config.json",
"quantize_config.json",
]
@classmethod
def from_config(cls, config: dict[str, Any]) -> "IPEXConfig":
method = cls.get_from_keys(config, ["quant_method"]).lower()
if method == "awq":
weight_bits = cls.get_from_keys(config, ["w_bit", "bits"])
group_size = cls.get_from_keys(config, ["q_group_size", "group_size"])
modules_to_not_convert = cls.get_from_keys_or(
config, ["modules_to_not_convert"], None
)
is_sym = not cls.get_from_keys_or(config, ["zero_point"], default=False)
return cls(
method,
weight_bits,
group_size,
modules_to_not_convert,
False,
False,
is_sym,
)
# otherwise for gptq
weight_bits = cls.get_from_keys(config, ["bits"])
group_size = cls.get_from_keys(config, ["group_size"])
lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
desc_act = cls.get_from_keys_or(config, ["desc_act"], default=False)
is_sym = cls.get_from_keys_or(config, ["sym"], default=True)
return cls(
method, weight_bits, group_size, [], desc_act, lm_head_quantized, is_sym
)
@classmethod
def override_quantization_method(
cls, hf_quant_cfg, user_quant
) -> QuantizationMethods | None:
if not current_platform.is_xpu():
return None
quant_method = hf_quant_cfg.get("quant_method", "").lower()
if quant_method in ["awq", "gptq"]:
return cls.get_name()
return None
def get_quant_method(
self, layer: torch.nn.Module, prefix: str
) -> "LinearMethodBase | None":
if isinstance(layer, LinearBase):
if self.method == "awq":
if is_layer_skipped(
prefix,
self.modules_to_not_convert,
self.packed_modules_mapping,
skip_with_substr=True,
):
return UnquantizedLinearMethod()
return IPEXAWQLinearMethod(self)
if self.method == "gptq":
return IPEXGPTQLinearMethod(self)
return None
class IPEXGPTQLinearMethod(GPTQLinearMethod):
"""GPTQ linear method using IPEX for the CPU/XPU backend."""
def __init__(self, quant_config: IPEXConfig):
self.quant_config = quant_config # type: ignore
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
bias = layer.bias if not layer.skip_bias_add else None
try:
import intel_extension_for_pytorch as ipex
if version.parse(ipex.__version__) < version.parse(MIN_IPEX_VERSION):
raise ImportError(
"intel_extension_for_pytorch version is "
"wrong. Please install "
f"intel_extension_for_pytorch>={MIN_IPEX_VERSION}."
)
except ImportError as err:
raise ImportError(
"Please install "
f"intel_extension_for_pytorch>={MIN_IPEX_VERSION} via "
f"`pip install intel_extension_for_pytorch>={MIN_IPEX_VERSION}`"
" to use IPEX-AWQ linear method."
) from err
# Using the compute dtype (lowp_mode) as INT8 to leverage instructions
# with better performance.
lowp_mode = ipex.quantization.WoqLowpMode.INT8
# The weight will be de-packed from INT4 to INT8.
weight_dtype = ipex.quantization.WoqWeightDtype.INT4
# The float activation will be quantized (dynamic, per-token) to INT8.
act_quant_mode = ipex.quantization.WoqActQuantMode.PER_BATCH_IC_BLOCK
assert isinstance(self.quant_config, IPEXConfig)
qconfig = ipex.quantization.get_weight_only_quant_qconfig_mapping(
weight_dtype=weight_dtype,
lowp_mode=lowp_mode,
act_quant_mode=act_quant_mode,
group_size=self.quant_config.group_size,
)
layer.ipex_output_size = layer.qweight.shape[-1]
g_idx = layer.g_idx if self.quant_config.desc_act else None
layer.ipex_qlinear = (
ipex.llm.quantization.woq_linear.IPEXWeightOnlyQuantizedLinear.from_weight(
layer.qweight,
layer.scales,
layer.qzeros,
layer.qweight.size(0),
layer.ipex_output_size,
qconfig=qconfig,
g_idx=g_idx,
bias=bias,
group_size=self.quant_config.group_size,
quant_method=IPEXConfig.IPEX_QUANT_METHOD_MAP["gptq"],
weight_qscheme="sym" if self.quant_config.is_sym else "asym",
)
)
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
reshaped_x = x.reshape(-1, x.shape[-1])
out = layer.ipex_qlinear(reshaped_x)
return out.reshape(x.shape[:-1] + (layer.ipex_output_size,))
class IPEXAWQLinearMethod(AWQLinearMethod):
"""AWQ linear method using IPEX for the CPU/XPU backend."""
def __init__(self, quant_config: IPEXConfig):
self.quant_config = quant_config # type: ignore
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
super().process_weights_after_loading(layer=layer)
bias = layer.bias if not layer.skip_bias_add else None
try:
import intel_extension_for_pytorch as ipex
if version.parse(ipex.__version__) < version.parse(MIN_IPEX_VERSION):
raise ImportError(
"intel_extension_for_pytorch version is "
"wrong. Please install "
f"intel_extension_for_pytorch>={MIN_IPEX_VERSION}."
)
except ImportError as err:
raise ImportError(
"Please install "
f"intel_extension_for_pytorch>={MIN_IPEX_VERSION} via "
f"`pip install intel_extension_for_pytorch>={MIN_IPEX_VERSION}`"
" to use IPEX-AWQ linear method."
) from err
# Using the compute dtype (lowp_mode) as INT8 to leverage instructions
# with better performance.
lowp_mode = ipex.quantization.WoqLowpMode.INT8
# The weight will be de-packed from INT4 to INT8.
weight_dtype = ipex.quantization.WoqWeightDtype.INT4
# The float activation will be quantized (dynamic, per-token) to INT8.
act_quant_mode = ipex.quantization.WoqActQuantMode.PER_BATCH
assert isinstance(self.quant_config, IPEXConfig)
qconfig = ipex.quantization.get_weight_only_quant_qconfig_mapping(
weight_dtype=weight_dtype,
lowp_mode=lowp_mode,
act_quant_mode=act_quant_mode,
group_size=self.quant_config.group_size,
)
layer.ipex_output_size = layer.qweight.size(1) * self.quant_config.pack_factor
layer.ipex_qlinear = (
ipex.llm.quantization.woq_linear.IPEXWeightOnlyQuantizedLinear.from_weight(
layer.qweight,
layer.scales,
layer.qzeros,
layer.qweight.size(0),
layer.ipex_output_size,
qconfig=qconfig,
bias=bias,
group_size=self.quant_config.group_size,
quant_method=IPEXConfig.IPEX_QUANT_METHOD_MAP["awq"], # type: ignore
weight_qscheme="sym" if self.quant_config.is_sym else "asym",
)
)
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
reshaped_x = x.reshape(-1, x.shape[-1])
out = layer.ipex_qlinear(reshaped_x)
return out.reshape(x.shape[:-1] + (layer.ipex_output_size,))
class XPUFp8LinearMethod(Fp8LinearMethod):
def __init__(self, quant_config: Fp8Config):
super().__init__(quant_config)
def process_weights_after_loading(self, layer: Module) -> None:
if getattr(layer, "_already_called_process_weights_after_loading", False):
return
# If checkpoint not serialized fp8, quantize the weights.
if not self.quant_config.is_checkpoint_fp8_serialized:
qweight, weight_scale = ops.scaled_fp8_quant(layer.weight, scale=None)
# Update the layer with the new values.
replace_parameter(layer, "weight", qweight.data)
replace_parameter(layer, "weight_scale", weight_scale.data)
layer.input_scale = None
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
weight = layer.weight.data
weight_scale = layer.weight_scale.data
output = torch.ops.torch_ipex.fp8_gemm_w8a16(
x, weight, True, weight_scale, bias
)
return output
class XPUFp8MoEMethod(Fp8OnlineMoEMethod):
def __init__(self, quant_config: Fp8Config, layer: torch.nn.Module):
super().__init__(quant_config, layer)
self.quant_config = quant_config
def process_weights_after_loading(self, layer: Module) -> None:
if getattr(layer, "_already_called_process_weights_after_loading", False):
return
if not self.quant_config.is_checkpoint_fp8_serialized:
fp8_dtype = current_platform.fp8_dtype()
w13_weight = torch.empty_like(layer.w13_weight.data, dtype=fp8_dtype)
w2_weight = torch.empty_like(layer.w2_weight.data, dtype=fp8_dtype)
# Re-initialize w13_scale because we directly quantize
# merged w13 weights and generate a single scaling factor.
layer.w13_weight_scale = torch.nn.Parameter(
torch.ones(
layer.local_num_experts,
dtype=torch.float32,
device=w13_weight.device,
),
requires_grad=False,
)
for expert in range(layer.local_num_experts):
w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
)
w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
ops.scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
)
replace_parameter(layer, "w13_weight", w13_weight)
replace_parameter(layer, "w2_weight", w2_weight)
import intel_extension_for_pytorch as ipex
ep_rank_start = self.moe.ep_rank * self.moe.num_local_experts
layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
layer.w13_weight,
layer.w2_weight,
w1_scale_inv=layer.w13_weight_scale,
w2_scale_inv=layer.w2_weight_scale,
a1_scale_inv=layer.w13_input_scale,
a2_scale_inv=layer.w2_input_scale,
use_prepack=True,
experts_start_id=ep_rank_start,
)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
return None
@property
def is_monolithic(self) -> bool:
return True
def apply_monolithic(
self,
layer: torch.nn.Module,
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor:
return layer.ipex_fusion(
x,
layer.use_grouped_topk,
layer.top_k,
router_logits,
layer.renormalize,
layer.topk_group,
layer.num_expert_group,
custom_routing_function=layer.custom_routing_function,
)
vllm/model_executor/layers/rotary_embedding/base.py
View file @ e1060448
...@@ -232,17 +232,14 @@ class RotaryEmbedding(RotaryEmbeddingBase): ...@@ -232,17 +232,14 @@ class RotaryEmbedding(RotaryEmbeddingBase):
query: torch.Tensor, query: torch.Tensor,
key: torch.Tensor | None = None, key: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]: ) -> tuple[torch.Tensor, torch.Tensor | None]:
from vllm._ipex_ops import ipex_ops as ops
self._match_cos_sin_cache_dtype(query) self._match_cos_sin_cache_dtype(query)
# ops.rotary_embedding() is an in-place operation # ops.rotary_embedding() is an in-place operation
# that updates the query and key tensors. # that updates the query and key tensors.
if key is None: if key is None:
# XPU kernel doesn't support key=None so fall back to native impl
# TODO(sarckk): add support for optional key in
# ipex.llm.functional.rotary_embedding_batched
return self.forward_native(positions, query, key) return self.forward_native(positions, query, key)
else: else:
from vllm import _custom_ops as ops
ops.rotary_embedding( ops.rotary_embedding(
positions, positions,
query, query,
......
vllm/platforms/__init__.py
View file @ e1060448
...@@ -132,8 +132,6 @@ def xpu_platform_plugin() -> str | None: ...@@ -132,8 +132,6 @@ def xpu_platform_plugin() -> str | None:
is_xpu = False is_xpu = False
logger.debug("Checking if XPU platform is available.") logger.debug("Checking if XPU platform is available.")
try: try:
# installed IPEX if the machine has XPUs.
import intel_extension_for_pytorch # noqa: F401
import torch import torch
if supports_xccl(): if supports_xccl():
......
vllm/platforms/xpu.py
View file @ e1060448
...@@ -7,6 +7,11 @@ from typing import TYPE_CHECKING ...@@ -7,6 +7,11 @@ from typing import TYPE_CHECKING
import torch import torch
# import custom ops, trigger op registration
import vllm_xpu_kernels._C # noqa
import vllm_xpu_kernels._moe_C # noqa
import vllm_xpu_kernels._xpu_C # noqa
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.v1.attention.backends.registry import AttentionBackendEnum from vllm.v1.attention.backends.registry import AttentionBackendEnum
...@@ -55,6 +60,9 @@ class XPUPlatform(Platform): ...@@ -55,6 +60,9 @@ class XPUPlatform(Platform):
dtype = attn_selector_config.dtype dtype = attn_selector_config.dtype
if attn_selector_config.use_sparse: if attn_selector_config.use_sparse:
raise NotImplementedError("Sparse Attention is not supported on XPU.") raise NotImplementedError("Sparse Attention is not supported on XPU.")
if attn_selector_config.use_mla:
logger.info_once("Using Triton MLA backend on V1 engine.")
return AttentionBackendEnum.TRITON_MLA.get_path()
if selected_backend == AttentionBackendEnum.TRITON_ATTN: if selected_backend == AttentionBackendEnum.TRITON_ATTN:
logger.info_once("Using Triton backend.") logger.info_once("Using Triton backend.")
return AttentionBackendEnum.TRITON_ATTN.get_path() return AttentionBackendEnum.TRITON_ATTN.get_path()
...@@ -78,9 +86,9 @@ class XPUPlatform(Platform): ...@@ -78,9 +86,9 @@ class XPUPlatform(Platform):
@classmethod @classmethod
def get_supported_vit_attn_backends(cls) -> list["AttentionBackendEnum"]: def get_supported_vit_attn_backends(cls) -> list["AttentionBackendEnum"]:
# XPU only supports FLASH_ATTN for vision attention.
return [ return [
AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
] ]
@classmethod @classmethod
...@@ -145,7 +153,7 @@ class XPUPlatform(Platform): ...@@ -145,7 +153,7 @@ class XPUPlatform(Platform):
def check_and_update_config(cls, vllm_config: VllmConfig) -> None: def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
cache_config = vllm_config.cache_config cache_config = vllm_config.cache_config
model_config = vllm_config.model_config model_config = vllm_config.model_config
# in V1(or with ipex chunked prefill) block_size is 64 # in V1(or with chunked prefill) block_size is 64
if cache_config and cache_config.block_size is None: if cache_config and cache_config.block_size is None:
cache_config.block_size = 64 cache_config.block_size = 64
...@@ -206,7 +214,7 @@ class XPUPlatform(Platform): ...@@ -206,7 +214,7 @@ class XPUPlatform(Platform):
@classmethod @classmethod
def fp8_dtype(cls) -> torch.dtype: def fp8_dtype(cls) -> torch.dtype:
return torch.float8_e5m2 return torch.float8_e4m3fn
@classmethod @classmethod
def is_data_center_gpu(cls) -> bool: def is_data_center_gpu(cls) -> bool:
......
vllm/v1/attention/backends/fa_utils.py
View file @ e1060448
...@@ -16,12 +16,13 @@ if current_platform.is_cuda(): ...@@ -16,12 +16,13 @@ if current_platform.is_cuda():
) )
elif current_platform.is_xpu(): elif current_platform.is_xpu():
from vllm import _custom_ops as ops
reshape_and_cache_flash = ops.reshape_and_cache_flash
from vllm._ipex_ops import ipex_ops from vllm._ipex_ops import ipex_ops
reshape_and_cache_flash = ipex_ops.reshape_and_cache_flash
flash_attn_varlen_func = ipex_ops.flash_attn_varlen_func # type: ignore[assignment] flash_attn_varlen_func = ipex_ops.flash_attn_varlen_func # type: ignore[assignment]
get_scheduler_metadata = ipex_ops.get_scheduler_metadata # type: ignore[assignment] get_scheduler_metadata = ipex_ops.get_scheduler_metadata # type: ignore[assignment]
elif current_platform.is_rocm(): elif current_platform.is_rocm():
try: try:
from flash_attn import flash_attn_varlen_func # type: ignore[no-redef] from flash_attn import flash_attn_varlen_func # type: ignore[no-redef]
......
vllm/v1/attention/backends/registry.py
View file @ e1060448
...@@ -69,7 +69,6 @@ class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta): ...@@ -69,7 +69,6 @@ class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
"vllm.v1.attention.backends.mla.flashmla_sparse.FlashMLASparseBackend" "vllm.v1.attention.backends.mla.flashmla_sparse.FlashMLASparseBackend"
) )
FLASH_ATTN_MLA = "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend" FLASH_ATTN_MLA = "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend"
IPEX = "vllm.v1.attention.backends.ipex.IpexAttentionBackend"
NO_ATTENTION = "vllm.v1.attention.backends.no_attention.NoAttentionBackend" NO_ATTENTION = "vllm.v1.attention.backends.no_attention.NoAttentionBackend"
FLEX_ATTENTION = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend" FLEX_ATTENTION = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend"
TREE_ATTN = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend" TREE_ATTN = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend"
......
vllm/v1/worker/xpu_worker.py
View file @ e1060448
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import gc
import os import os
from typing import Any from typing import Any
import torch import torch
import torch.distributed
from vllm.config import VllmConfig from vllm.config import VllmConfig
from vllm.logger import init_logger from vllm.logger import init_logger
...@@ -85,7 +85,14 @@ class XPUWorker(Worker): ...@@ -85,7 +85,14 @@ class XPUWorker(Worker):
current_platform.dist_backend, current_platform.dist_backend,
) )
# Set random seed.
set_random_seed(self.model_config.seed)
# Now take memory snapshot after NCCL is initialized
gc.collect()
torch.xpu.empty_cache() torch.xpu.empty_cache()
# take current memory snapshot
self.init_snapshot = init_snapshot = MemorySnapshot(device=self.device) self.init_snapshot = init_snapshot = MemorySnapshot(device=self.device)
self.requested_memory = request_memory(init_snapshot, self.cache_config) self.requested_memory = request_memory(init_snapshot, self.cache_config)
logger.debug("worker init memory snapshot: %r", self.init_snapshot) logger.debug("worker init memory snapshot: %r", self.init_snapshot)
...@@ -93,9 +100,6 @@ class XPUWorker(Worker): ...@@ -93,9 +100,6 @@ class XPUWorker(Worker):
"worker requested memory: %sGiB", format_gib(self.requested_memory) "worker requested memory: %sGiB", format_gib(self.requested_memory)
) )
# Set random seed.
set_random_seed(self.model_config.seed)
# Initialize workspace manager # Initialize workspace manager
num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1 num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1
init_workspace_manager(self.device, num_ubatches) init_workspace_manager(self.device, num_ubatches)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment