Skip to content

GitLab

Commit c721b814 authored by zhuwenwen's avatar zhuwenwen
Browse files

sync v0.15.1

parent d53fe7e5
Hide whitespace changes
Inline Side-by-side
Showing with 31 additions and 2330 deletions
vllm/envs.py
View file @ c721b814
...@@ -87,7 +87,6 @@ if TYPE_CHECKING: ...@@ -87,7 +87,6 @@ if TYPE_CHECKING:
VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5 # seconds VLLM_HTTP_TIMEOUT_KEEP_ALIVE: int = 5 # seconds
VLLM_PLUGINS: list[str] | None = None VLLM_PLUGINS: list[str] | None = None
VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None VLLM_LORA_RESOLVER_CACHE_DIR: str | None = None
VLLM_LORA_RESOLVER_HF_REPO_LIST: str | None = None
# Deprecated env variables for profiling, kept for backward compatibility # Deprecated env variables for profiling, kept for backward compatibility
# See also vllm/config/profiler.py and `--profiler-config` argument # See also vllm/config/profiler.py and `--profiler-config` argument
VLLM_TORCH_CUDA_PROFILE: str | None = None VLLM_TORCH_CUDA_PROFILE: str | None = None
...@@ -289,11 +288,16 @@ def use_aot_compile() -> bool: ...@@ -289,11 +288,16 @@ def use_aot_compile() -> bool:
from vllm.model_executor.layers.batch_invariant import ( from vllm.model_executor.layers.batch_invariant import (
vllm_is_batch_invariant, vllm_is_batch_invariant,
) )
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_torch_equal_or_newer from vllm.utils.torch_utils import is_torch_equal_or_newer
default_value = ( default_value = (
"1" "1"
if is_torch_equal_or_newer("2.10.0.dev") and not disable_compile_cache() if is_torch_equal_or_newer("2.10.0.dev")
and not disable_compile_cache()
# Disabling AOT_COMPILE for CPU
# See: https://github.com/vllm-project/vllm/issues/32033
and not current_platform.is_cpu()
else "0" else "0"
) )
...@@ -870,13 +874,6 @@ environment_variables: dict[str, Callable[[], Any]] = { ...@@ -870,13 +874,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv( "VLLM_LORA_RESOLVER_CACHE_DIR": lambda: os.getenv(
"VLLM_LORA_RESOLVER_CACHE_DIR", None "VLLM_LORA_RESOLVER_CACHE_DIR", None
), ),
# A remote HF repo(s) containing one or more LoRA adapters, which
# may be downloaded and leveraged as needed. Only works if plugins
# are enabled and VLLM_ALLOW_RUNTIME_LORA_UPDATING is enabled.
# Values should be comma separated.
"VLLM_LORA_RESOLVER_HF_REPO_LIST": lambda: os.getenv(
"VLLM_LORA_RESOLVER_HF_REPO_LIST", None
),
# Enables torch CUDA profiling if set to 1. # Enables torch CUDA profiling if set to 1.
# Deprecated, see profiler_config. # Deprecated, see profiler_config.
"VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"), "VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"),
...@@ -884,7 +881,7 @@ environment_variables: dict[str, Callable[[], Any]] = { ...@@ -884,7 +881,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Deprecated, see profiler_config. # Deprecated, see profiler_config.
"VLLM_TORCH_PROFILER_DIR": lambda: os.getenv("VLLM_TORCH_PROFILER_DIR"), "VLLM_TORCH_PROFILER_DIR": lambda: os.getenv("VLLM_TORCH_PROFILER_DIR"),
# Enable torch profiler to record shapes if set to 1. # Enable torch profiler to record shapes if set to 1.
# Deprecated, see profiler_config. # Deprecated, see profiler_config.
"VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: ( "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: (
os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES") os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES")
), ),
......
vllm/logging_utils/__init__.py
View file @ c721b814
# 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 vllm.logging_utils.access_log_filter import (
UvicornAccessLogFilter,
create_uvicorn_log_config,
)
from vllm.logging_utils.formatter import ColoredFormatter, NewLineFormatter from vllm.logging_utils.formatter import ColoredFormatter, NewLineFormatter
from vllm.logging_utils.lazy import lazy from vllm.logging_utils.lazy import lazy
from vllm.logging_utils.log_time import logtime from vllm.logging_utils.log_time import logtime
...@@ -12,8 +8,6 @@ from vllm.logging_utils.log_time import logtime ...@@ -12,8 +8,6 @@ from vllm.logging_utils.log_time import logtime
__all__ = [ __all__ = [
"NewLineFormatter", "NewLineFormatter",
"ColoredFormatter", "ColoredFormatter",
"UvicornAccessLogFilter",
"create_uvicorn_log_config",
"lazy", "lazy",
"logtime", "logtime",
] ]
vllm/logging_utils/access_log_filter.py deleted 100644 → 0
View file @ d53fe7e5
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Access log filter for uvicorn to exclude specific endpoints from logging.
This module provides a logging filter that can be used to suppress access logs
for specific endpoints (e.g., /health, /metrics) to reduce log noise in
production environments.
"""
import logging
from urllib.parse import urlparse
class UvicornAccessLogFilter(logging.Filter):
"""
A logging filter that excludes access logs for specified endpoint paths.
This filter is designed to work with uvicorn's access logger. It checks
the log record's arguments for the request path and filters out records
matching the excluded paths.
Uvicorn access log format:
'%s - "%s %s HTTP/%s" %d'
(client_addr, method, path, http_version, status_code)
Example:
127.0.0.1:12345 - "GET /health HTTP/1.1" 200
Args:
excluded_paths: A list of URL paths to exclude from logging.
Paths are matched exactly.
Example: ["/health", "/metrics"]
"""
def __init__(self, excluded_paths: list[str] | None = None):
super().__init__()
self.excluded_paths = set(excluded_paths or [])
def filter(self, record: logging.LogRecord) -> bool:
"""
Determine if the log record should be logged.
Args:
record: The log record to evaluate.
Returns:
True if the record should be logged, False otherwise.
"""
if not self.excluded_paths:
return True
# This filter is specific to uvicorn's access logs.
if record.name != "uvicorn.access":
return True
# The path is the 3rd argument in the log record's args tuple.
# See uvicorn's access logging implementation for details.
log_args = record.args
if isinstance(log_args, tuple) and len(log_args) >= 3:
path_with_query = log_args[2]
# Get path component without query string.
if isinstance(path_with_query, str):
path = urlparse(path_with_query).path
if path in self.excluded_paths:
return False
return True
def create_uvicorn_log_config(
excluded_paths: list[str] | None = None,
log_level: str = "info",
) -> dict:
"""
Create a uvicorn logging configuration with access log filtering.
This function generates a logging configuration dictionary that can be
passed to uvicorn's `log_config` parameter. It sets up the access log
filter to exclude specified paths.
Args:
excluded_paths: List of URL paths to exclude from access logs.
log_level: The log level for uvicorn loggers.
Returns:
A dictionary containing the logging configuration.
Example:
>>> config = create_uvicorn_log_config(["/health", "/metrics"])
>>> uvicorn.run(app, log_config=config)
"""
config = {
"version": 1,
"disable_existing_loggers": False,
"filters": {
"access_log_filter": {
"()": UvicornAccessLogFilter,
"excluded_paths": excluded_paths or [],
},
},
"formatters": {
"default": {
"()": "uvicorn.logging.DefaultFormatter",
"fmt": "%(levelprefix)s %(message)s",
"use_colors": None,
},
"access": {
"()": "uvicorn.logging.AccessFormatter",
"fmt": '%(levelprefix)s %(client_addr)s - "%(request_line)s" %(status_code)s', # noqa: E501
},
},
"handlers": {
"default": {
"formatter": "default",
"class": "logging.StreamHandler",
"stream": "ext://sys.stderr",
},
"access": {
"formatter": "access",
"class": "logging.StreamHandler",
"stream": "ext://sys.stdout",
"filters": ["access_log_filter"],
},
},
"loggers": {
"uvicorn": {
"handlers": ["default"],
"level": log_level.upper(),
"propagate": False,
},
"uvicorn.error": {
"level": log_level.upper(),
"handlers": ["default"],
"propagate": False,
},
"uvicorn.access": {
"handlers": ["access"],
"level": log_level.upper(),
"propagate": False,
},
},
}
return config
vllm/lora/ops/triton_ops/fused_moe_lora_op.py
View file @ c721b814
...@@ -62,7 +62,6 @@ def _fused_moe_lora_kernel( ...@@ -62,7 +62,6 @@ def _fused_moe_lora_kernel(
num_experts, num_experts,
lora_ids, lora_ids,
adapter_enabled, adapter_enabled,
max_loras, # <<< PR2: rename, used for masks when grid axis-2 != max_loras
# The stride variables represent how much to increase the ptr by when # The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is # moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down # how much to increase `a_ptr` by to get the element one row down
...@@ -84,7 +83,6 @@ def _fused_moe_lora_kernel( ...@@ -84,7 +83,6 @@ def _fused_moe_lora_kernel(
num_slice_c: tl.constexpr, num_slice_c: tl.constexpr,
top_k: tl.constexpr, top_k: tl.constexpr,
MUL_ROUTED_WEIGHT: tl.constexpr, MUL_ROUTED_WEIGHT: tl.constexpr,
USE_B_L2_CACHE: tl.constexpr, # new, enable .ca load for B
BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
...@@ -106,13 +104,10 @@ def _fused_moe_lora_kernel( ...@@ -106,13 +104,10 @@ def _fused_moe_lora_kernel(
if moe_enabled == 0: if moe_enabled == 0:
# Early exit for the no moe lora case. # Early exit for the no moe lora case.
return return
# The grid's axis-2 dimension is max_loras + 1 to accommodate the -1 sentinel. # The grid size on axis 2 is (max_loras + 1) to handle the no-lora case
# This guard ensures we don't access sorted_token_ids / expert_ids / # (lora_id == -1), but sorted_token_ids and expert_ids are allocated with
# num_tokens_post_padded beyond their allocated bounds if an invalid # shape (max_loras, ...). Use (num_programs - 1) for correct bounds checking.
# lora_id somehow appears. Although the caller should pass correct max_loras = tl.num_programs(axis=2) - 1
# max_loras, defensive programming prevents accidental out-of-bounds.
if lora_id >= max_loras:
return
grid_k = tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K) grid_k = tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)
# calculate pid_m,pid_n # calculate pid_m,pid_n
...@@ -141,11 +136,10 @@ def _fused_moe_lora_kernel( ...@@ -141,11 +136,10 @@ def _fused_moe_lora_kernel(
cur_b_ptr = tl.load(b_ptr + slice_id).to(tl.pointer_type(c_ptr.dtype.element_ty)) cur_b_ptr = tl.load(b_ptr + slice_id).to(tl.pointer_type(c_ptr.dtype.element_ty))
cur_c_ptr = c_ptr + (slice_id % num_slice_c) * slice_c_size cur_c_ptr = c_ptr + (slice_id % num_slice_c) * slice_c_size
# remove modulo wrap-around offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
offs_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int32)
offs_k = pid_sk * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K) offs_k = pid_sk * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int32) offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
token_ind = stride_tl * lora_id + offs_token_id token_ind = stride_tl * lora_id + offs_token_id
offs_token = tl.load( offs_token = tl.load(
sorted_token_ids_ptr + token_ind, sorted_token_ids_ptr + token_ind,
...@@ -183,11 +177,7 @@ def _fused_moe_lora_kernel( ...@@ -183,11 +177,7 @@ def _fused_moe_lora_kernel(
# before continuing. # before continuing.
# pre-fetch lora weight # pre-fetch lora weight
# add (offs_bn < N) mask; optional .ca for B # add (offs_bn < N) mask; optional .ca for B
b_mask = (offs_k[:, None] < k_remaining) & (offs_bn[None, :] < N) b = tl.load(b_ptrs, mask=offs_k[:, None] < k_remaining, other=0.0)
if USE_B_L2_CACHE:
b = tl.load(b_ptrs, mask=b_mask, other=0.0, cache_modifier=".ca")
else:
b = tl.load(b_ptrs, mask=b_mask, other=0.0)
if USE_GDC and not IS_PRIMARY: if USE_GDC and not IS_PRIMARY:
tl.extra.cuda.gdc_wait() tl.extra.cuda.gdc_wait()
...@@ -288,7 +278,6 @@ def _fused_moe_lora_shrink( ...@@ -288,7 +278,6 @@ def _fused_moe_lora_shrink(
num_experts, num_experts,
lora_ids, lora_ids,
adapter_enabled, adapter_enabled,
lora_a_stacked[0].shape[0],
qcurr_hidden_states.stride(0), qcurr_hidden_states.stride(0),
qcurr_hidden_states.stride(1), qcurr_hidden_states.stride(1),
w1_lora_a_stacked.stride(0), w1_lora_a_stacked.stride(0),
...@@ -305,7 +294,6 @@ def _fused_moe_lora_shrink( ...@@ -305,7 +294,6 @@ def _fused_moe_lora_shrink(
num_slice_c=num_slices, num_slice_c=num_slices,
top_k=1 if mul_routed_weight else top_k_num, top_k=1 if mul_routed_weight else top_k_num,
MUL_ROUTED_WEIGHT=False, MUL_ROUTED_WEIGHT=False,
USE_B_L2_CACHE=True, # new
IS_PRIMARY=True, IS_PRIMARY=True,
**shrink_config, **shrink_config,
) )
...@@ -391,7 +379,6 @@ def _fused_moe_lora_expand( ...@@ -391,7 +379,6 @@ def _fused_moe_lora_expand(
num_experts, num_experts,
lora_ids, lora_ids,
adapter_enabled, adapter_enabled,
lora_b_stacked[0].shape[0],
a_intermediate_cache1.stride(0), a_intermediate_cache1.stride(0),
a_intermediate_cache1.stride(1), a_intermediate_cache1.stride(1),
w1_lora_b_stacked.stride(0), w1_lora_b_stacked.stride(0),
...@@ -408,7 +395,6 @@ def _fused_moe_lora_expand( ...@@ -408,7 +395,6 @@ def _fused_moe_lora_expand(
num_slice_c=num_slices, num_slice_c=num_slices,
top_k=1, top_k=1,
MUL_ROUTED_WEIGHT=mul_routed_weight, MUL_ROUTED_WEIGHT=mul_routed_weight,
USE_B_L2_CACHE=True, # new
IS_PRIMARY=False, IS_PRIMARY=False,
**expand_config, **expand_config,
) )
......
vllm/model_executor/layers/attention/mla_attention.py
View file @ c721b814
...@@ -1554,7 +1554,6 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]): ...@@ -1554,7 +1554,6 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v self, prefill: MLACommonPrefillMetadata, chunk_idx: int, q, k, v
): ):
assert isinstance(prefill, FlashInferPrefillMetadata) assert isinstance(prefill, FlashInferPrefillMetadata)
attn_out, lse = prefill.prefill_chunks[chunk_idx].run( attn_out, lse = prefill.prefill_chunks[chunk_idx].run(
q=q, q=q,
k=k, k=k,
......
vllm/model_executor/layers/fused_moe/all2all_utils.py
View file @ c721b814
...@@ -7,27 +7,18 @@ import torch ...@@ -7,27 +7,18 @@ import torch
from vllm.distributed import ( from vllm.distributed import (
get_ep_group, get_ep_group,
) )
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe.config import ( from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig, FusedMoEConfig,
FusedMoEParallelConfig, FusedMoEParallelConfig,
FusedMoEQuantConfig, FusedMoEQuantConfig,
) )
from vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize import (
FlashInferA2APrepareAndFinalize,
)
from vllm.model_executor.layers.fused_moe.modular_kernel import ( from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEPrepareAndFinalize, FusedMoEPrepareAndFinalize,
) )
from vllm.model_executor.layers.fused_moe.prepare_finalize import (
MoEPrepareAndFinalizeNaiveEP,
MoEPrepareAndFinalizeNoEP,
)
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.utils.import_utils import has_deep_ep, has_mori, has_pplx from vllm.utils.import_utils import has_deep_ep, has_mori, has_pplx
logger = init_logger(__name__)
if current_platform.is_cuda_alike(): if current_platform.is_cuda_alike():
if has_pplx(): if has_pplx():
from .pplx_prepare_finalize import ( from .pplx_prepare_finalize import (
...@@ -80,45 +71,19 @@ def maybe_make_prepare_finalize( ...@@ -80,45 +71,19 @@ def maybe_make_prepare_finalize(
moe: FusedMoEConfig, moe: FusedMoEConfig,
quant_config: FusedMoEQuantConfig | None, quant_config: FusedMoEQuantConfig | None,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None, routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
allow_new_interface: bool = False,
) -> FusedMoEPrepareAndFinalize | None: ) -> FusedMoEPrepareAndFinalize | None:
# NOTE(rob): we are migrating each quant_method to hold the MK
# in all cases. The allow_new_interface=False flag allow us to fall
# back to the old method for methods that have not yet been migrated.
#
# In old method:
# * maybe_init_modular_kernel() calls this function. If we are
# using no Dp/Ep or naive all2all, we return None this function
# returns None and no ModularKernelMethod is created. If non-naive
# all2all is used, this returns a PrepareAndFinalize object and
# a ModularKernelMethod is created.
# In new method:
# * maybe_make_prepare_finalize() is called from the oracle. We
# always return a PrepareAndFinalize object and the quant method
# holds the ModularKernel.
if not moe.moe_parallel_config.use_all2all_kernels: if not moe.moe_parallel_config.use_all2all_kernels:
if not allow_new_interface: return None
return None
# For DP/TP case, fall back to naive P/F.
if moe.moe_parallel_config.dp_size > 1:
logger.info_once(
"Detected DP deployment with no --enable-expert-parallel. "
"Falling back to AllGather+ReduceScatter dispatch/combine."
)
return MoEPrepareAndFinalizeNaiveEP(
is_sequence_parallel=moe.moe_parallel_config.is_sequence_parallel,
num_dispatchers=(
get_ep_group().device_communicator.all2all_manager.world_size
),
)
else:
return MoEPrepareAndFinalizeNoEP()
all2all_manager = get_ep_group().device_communicator.all2all_manager all2all_manager = get_ep_group().device_communicator.all2all_manager
assert all2all_manager is not None assert all2all_manager is not None
prepare_finalize: FusedMoEPrepareAndFinalize | None = None prepare_finalize: FusedMoEPrepareAndFinalize | None = None
# TODO(rob): update this as part of the MoE refactor.
assert not moe.use_flashinfer_cutlass_kernels, (
"Must be created in modelopt.py or fp8.py"
)
if moe.use_pplx_kernels: if moe.use_pplx_kernels:
assert quant_config is not None assert quant_config is not None
...@@ -239,16 +204,4 @@ def maybe_make_prepare_finalize( ...@@ -239,16 +204,4 @@ def maybe_make_prepare_finalize(
use_fp8_dispatch=use_fp8_dispatch, use_fp8_dispatch=use_fp8_dispatch,
) )
elif moe.use_fi_all2allv_kernels:
assert quant_config is not None
prepare_finalize = FlashInferA2APrepareAndFinalize(
num_dispatchers=all2all_manager.world_size,
)
elif moe.use_naive_all2all_kernels and allow_new_interface:
prepare_finalize = MoEPrepareAndFinalizeNaiveEP(
is_sequence_parallel=(moe.moe_parallel_config.is_sequence_parallel),
num_dispatchers=all2all_manager.world_size,
)
return prepare_finalize return prepare_finalize
vllm/model_executor/layers/fused_moe/config.py
View file @ c721b814
...@@ -20,6 +20,7 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import ( ...@@ -20,6 +20,7 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
) )
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
from vllm.utils.import_utils import has_triton_kernels from vllm.utils.import_utils import has_triton_kernels
from vllm.utils.math_utils import cdiv from vllm.utils.math_utils import cdiv
...@@ -861,7 +862,6 @@ class FusedMoEParallelConfig: ...@@ -861,7 +862,6 @@ class FusedMoEParallelConfig:
use_ep: bool # whether to use EP or not use_ep: bool # whether to use EP or not
all2all_backend: str # all2all backend for MoE communication all2all_backend: str # all2all backend for MoE communication
is_sequence_parallel: bool # whether sequence parallelism is used
enable_eplb: bool # whether to enable expert load balancing enable_eplb: bool # whether to enable expert load balancing
@property @property
...@@ -883,12 +883,6 @@ class FusedMoEParallelConfig: ...@@ -883,12 +883,6 @@ class FusedMoEParallelConfig:
def use_deepep_ll_kernels(self): def use_deepep_ll_kernels(self):
return self.use_all2all_kernels and self.all2all_backend == "deepep_low_latency" return self.use_all2all_kernels and self.all2all_backend == "deepep_low_latency"
@property
def use_fi_all2allv_kernels(self):
return (
self.use_all2all_kernels and self.all2all_backend == "flashinfer_all2allv"
)
@property @property
def use_batched_activation_format(self): def use_batched_activation_format(self):
return self.use_deepep_ll_kernels or self.use_pplx_kernels return self.use_deepep_ll_kernels or self.use_pplx_kernels
...@@ -1020,7 +1014,6 @@ class FusedMoEParallelConfig: ...@@ -1020,7 +1014,6 @@ class FusedMoEParallelConfig:
ep_rank=0, ep_rank=0,
use_ep=False, use_ep=False,
all2all_backend=vllm_parallel_config.all2all_backend, all2all_backend=vllm_parallel_config.all2all_backend,
is_sequence_parallel=vllm_parallel_config.use_sequence_parallel_moe,
enable_eplb=vllm_parallel_config.enable_eplb, enable_eplb=vllm_parallel_config.enable_eplb,
) )
# DP + EP / TP + EP / DP + TP + EP # DP + EP / TP + EP / DP + TP + EP
...@@ -1040,7 +1033,6 @@ class FusedMoEParallelConfig: ...@@ -1040,7 +1033,6 @@ class FusedMoEParallelConfig:
ep_rank=ep_rank, ep_rank=ep_rank,
use_ep=True, use_ep=True,
all2all_backend=vllm_parallel_config.all2all_backend, all2all_backend=vllm_parallel_config.all2all_backend,
is_sequence_parallel=vllm_parallel_config.use_sequence_parallel_moe,
enable_eplb=vllm_parallel_config.enable_eplb, enable_eplb=vllm_parallel_config.enable_eplb,
) )
...@@ -1059,7 +1051,6 @@ class FusedMoEParallelConfig: ...@@ -1059,7 +1051,6 @@ class FusedMoEParallelConfig:
use_ep=False, use_ep=False,
all2all_backend="naive", all2all_backend="naive",
enable_eplb=False, enable_eplb=False,
is_sequence_parallel=False,
) )
...@@ -1154,9 +1145,12 @@ class FusedMoEConfig: ...@@ -1154,9 +1145,12 @@ class FusedMoEConfig:
return self.moe_parallel_config.use_mori_kernels return self.moe_parallel_config.use_mori_kernels
@property @property
def use_fi_all2allv_kernels(self): def use_flashinfer_cutlass_kernels(self):
return self.moe_parallel_config.use_fi_all2allv_kernels """
Whether to use FlashInfer cutlass kernels for NVFP4 MoE.
@property """
def use_naive_all2all_kernels(self): return (
return self.moe_parallel_config.use_naive_all2all_kernels envs.VLLM_USE_FLASHINFER_MOE_FP4
and has_flashinfer_cutlass_fused_moe()
and envs.VLLM_FLASHINFER_MOE_BACKEND == "throughput"
)
vllm/model_executor/layers/fused_moe/configs/E=128,N=128,device_name=BW200.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
}
}
vllm/model_executor/layers/fused_moe/configs/E=128,N=128,device_name=BW200_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=BW200_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=DCU_K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=BW200.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"2048": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=BW200.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=BW200_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=DCU_K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=32,N=512,device_name=BW200_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=32,N=512,device_name=DCU_K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=32,N=512,device_name=K100_AI_nn.json deleted 100644 → 0
View file @ d53fe7e5
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
}
}
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