Skip to content

GitLab

Commit 7c4f76e3 authored by zhuwenwen's avatar zhuwenwen
Browse files

merge v0.4.0

parents 2da0dd3e 51c31bc1
Hide whitespace changes
Inline Side-by-side
Showing with 1639 additions and 515 deletions
vllm/lora/models.py
View file @ 7c4f76e3
import copy
import json
import logging
import math
import os
import re
from typing import (Any, Callable, Dict, Hashable, List, Optional, Tuple, Type)
from typing import Callable, Dict, Hashable, List, Optional, Tuple, Type
import safetensors.torch
import torch
from torch import nn
from vllm.config import LoRAConfig
from vllm.utils import LRUCache, in_wsl
from vllm.lora.layers import BaseLayerWithLoRA, LoRAMapping, from_layer, from_layer_sampler
from vllm.logger import init_logger
from vllm.lora.layers import (BaseLayerWithLoRA, LoRAMapping, from_layer,
from_layer_logits_processor)
from vllm.lora.lora import LoRALayerWeights, PackedLoRALayerWeights
from vllm.lora.utils import parse_fine_tuned_lora_name, replace_submodule
from vllm.utils import LRUCache, is_pin_memory_available
logger = logging.getLogger(__name__)
logger = init_logger(__name__)
_GLOBAL_LORA_ID = 0
......@@ -142,7 +142,7 @@ class LoRAModel:
embedding_padding_modules: Optional[List[str]] = None,
) -> "LoRAModel":
"""Create a LoRAModel from a dictionary of tensors."""
pin_memory = str(device) == "cpu" and not in_wsl()
pin_memory = str(device) == "cpu" and is_pin_memory_available()
loras: Dict[str, LoRALayerWeights] = {}
for tensor_name, tensor in tensors.items():
module_name, is_lora_a = parse_fine_tuned_lora_name(tensor_name)
......@@ -413,18 +413,22 @@ class LoRAModelManager:
for module_name, module in self.model.named_modules():
if not self._match_target_modules(module_name):
continue
parts = module_name.split(".")[-1]
packed_moduled_lst = self.packed_modules_mapping.get(parts, [])
new_module = replace_submodule(
self.model, module_name,
from_layer(module, self.lora_slots, self.lora_config,
self.model.config))
packed_moduled_lst, self.model.config))
# (yard1): TODO make this more robust
if "lm_head" in module_name:
sampler_module = self.model.get_submodule("sampler")
logits_processor_module = self.model.get_submodule(
"logits_processor")
new_module = replace_submodule(
self.model, "sampler",
from_layer_sampler(sampler_module, module, self.lora_slots,
self.lora_config, self.model.config))
self.model, "logits_processor",
from_layer_logits_processor(logits_processor_module,
module, self.lora_slots,
self.lora_config,
self.model.config))
self.register_module(module_name, new_module)
self._register_packed_modules(module_name)
new_module.set_mapping(self.base_indices, self.sampler_indices,
......@@ -507,8 +511,10 @@ class LoRAModelManager:
def _register_packed_modules(self, module_full_name: str) -> None:
parts = module_full_name.split(".")
module_name = parts[-1]
replacements = self.packed_modules_mapping.get(module_name)
if not replacements:
replacements = self.packed_modules_mapping.get(module_name, [])
# When replacements is less than or equal to 1, it indicates that this
# module is not a packed module.
if len(replacements) <= 1:
return
prefix = ".".join(parts[:-1])
self.packed_modules[module_full_name] = [
......@@ -534,14 +540,14 @@ class LoRAModelManager:
replacement_loras)
class LoRALRUCache(LRUCache):
class LoRALRUCache(LRUCache[LoRAModel]):
def __init__(self, capacity: int, deactivate_lora_fn: Callable[[Hashable],
None]):
super().__init__(capacity)
self.deactivate_lora_fn = deactivate_lora_fn
def _on_remove(self, key: Hashable, value: Any):
def _on_remove(self, key: Hashable, value: LoRAModel):
logger.debug(f"Removing LoRA. int id: {key}")
self.deactivate_lora_fn(key)
return super()._on_remove(key, value)
......
vllm/lora/utils.py
View file @ 7c4f76e3
import logging
from typing import Tuple
from torch import nn
logger = logging.getLogger(__name__)
from vllm.logger import init_logger
logger = init_logger(__name__)
def replace_submodule(model: nn.Module, module_name: str,
......
vllm/lora/worker_manager.py
View file @ 7c4f76e3
import logging
from abc import ABC, abstractmethod, abstractproperty
from typing import Any, Dict, List, Optional, Set, Type
import torch
from vllm.config import LoRAConfig
from vllm.logger import init_logger
from vllm.lora.layers import LoRAMapping
from vllm.lora.models import (LoRAModel, LoRAModelManager,
LRUCacheLoRAModelManager, create_lora_manager)
from vllm.lora.request import LoRARequest
from vllm.lora.layers import LoRAMapping
from vllm.config import LoRAConfig
logger = logging.getLogger(__name__)
logger = init_logger(__name__)
class AbstractWorkerLoRAManager(ABC):
......@@ -154,10 +154,9 @@ class WorkerLoRAManager(AbstractWorkerLoRAManager):
f"LoRA rank {lora.rank} is greater than max_lora_rank "
f"{self.lora_config.max_lora_rank}.")
if lora.extra_vocab_size > self.lora_config.lora_extra_vocab_size:
raise ValueError(
f"LoRA added vocab size {lora.extra_vocab_size} is greater than "
f"lora_extra_vocab_size {self.lora_config.lora_extra_vocab_size}."
)
raise ValueError(f"LoRA added vocab size {lora.extra_vocab_size} "
f"is greater than lora_extra_vocab_size "
f"{self.lora_config.lora_extra_vocab_size}.")
return lora
def add_dummy_lora(self, lora_request: LoRARequest, rank: int) -> bool:
......
vllm/model_executor/__init__.py
View file @ 7c4f76e3
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.utils import set_random_seed, get_model
from vllm.model_executor.utils import set_random_seed
__all__ = [
"InputMetadata",
"get_model",
"SamplingMetadata",
"set_random_seed",
]
vllm/model_executor/guided_decoding.py
View file @ 7c4f76e3
......@@ -5,18 +5,51 @@ from enum import Enum
from functools import lru_cache
from json import dumps as json_dumps
from re import escape as regex_escape
from typing import Union, Tuple
from typing import Tuple, Union
from pydantic import BaseModel
from transformers import PreTrainedTokenizerBase
from vllm.entrypoints.openai.protocol import CompletionRequest, ChatCompletionRequest
from vllm.model_executor.guided_logits_processors import JSONLogitsProcessor, RegexLogitsProcessor
from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
CompletionRequest)
from vllm.model_executor.guided_logits_processors import (CFGLogitsProcessor,
JSONLogitsProcessor,
RegexLogitsProcessor)
class GuidedDecodingMode(Enum):
JSON = "json"
REGEX = "regex"
CHOICE = "choice"
GRAMMAR = "grammar"
# https://github.com/outlines-dev/outlines/blob/main/outlines/grammars/json.lark
# the main difference is that we changed the start: value to
# start: object | array, so we are denying scalar values as the root of the
# JSON. Starting with scalars as the root seems to cause llama to generate
# without stop.
JSON_GRAMMAR = r"""
?start: object | array
?value: object
| array
| UNESCAPED_STRING
| SIGNED_NUMBER -> number
| "true" -> true
| "false" -> false
| "null" -> null
array : "[" [value ("," value)*] "]"
object : "{" [pair ("," pair)*] "}"
pair : UNESCAPED_STRING ":" value
%import common.UNESCAPED_STRING
%import common.SIGNED_NUMBER
%import common.WS
%ignore WS
"""
global_thread_pool = None # used for generating logits processor fsm
......@@ -55,9 +88,6 @@ def _get_guide_and_mode(
) -> Tuple[str, GuidedDecodingMode]:
if request.guided_json:
if not isinstance(request.guided_json, (str, dict, BaseModel)):
raise TypeError("JSON schema must be str, dict, or BaseModel")
json = request.guided_json
if isinstance(json, dict):
# turn dict into hashable string
......@@ -67,33 +97,33 @@ def _get_guide_and_mode(
# with the same fields will get hashed the same
json = str(json.__signature__)
return json, GuidedDecodingMode.JSON
elif request.guided_regex:
if not isinstance(request.guided_regex, str):
raise TypeError("Regex must be string")
return request.guided_regex, GuidedDecodingMode.REGEX
elif request.guided_choice:
if not isinstance(request.guided_choice, list):
raise TypeError("Choices must be a list")
# choice just uses regex
choices = [
regex_escape(str(choice)) for choice in request.guided_choice
]
choices_regex = "(" + "|".join(choices) + ")"
return choices_regex, GuidedDecodingMode.CHOICE
elif request.guided_grammar:
return request.guided_grammar, GuidedDecodingMode.GRAMMAR
elif (request.response_format is not None
and request.response_format.type == "json_object"):
return JSON_GRAMMAR, GuidedDecodingMode.GRAMMAR
else:
return None, None
@lru_cache(maxsize=32)
def _get_cached_logits_processor(guide: str, tokenizer,
def _get_cached_logits_processor(guide: str,
tokenizer: PreTrainedTokenizerBase,
mode: GuidedDecodingMode):
if mode == GuidedDecodingMode.JSON:
return JSONLogitsProcessor(guide, tokenizer)
elif mode == GuidedDecodingMode.REGEX or mode == GuidedDecodingMode.CHOICE:
return RegexLogitsProcessor(guide, tokenizer)
elif mode == GuidedDecodingMode.GRAMMAR:
return CFGLogitsProcessor(guide, tokenizer)
else:
raise ValueError(f"Unknown guided decoding mode {mode}")
vllm/model_executor/guided_logits_processors.py
View file @ 7c4f76e3
......@@ -16,30 +16,60 @@
import json
import math
from collections import defaultdict
from typing import Union, DefaultDict, Dict, List, Optional
from typing import Callable, DefaultDict, Dict, List, Optional, Union
import torch
from pydantic import BaseModel
from outlines.fsm.fsm import RegexFSM
from outlines.fsm.fsm import CFGFSM, RegexFSM
from outlines.fsm.json_schema import build_regex_from_schema
from pydantic import BaseModel
from transformers import PreTrainedTokenizerBase
class RegexLogitsProcessor:
class BaseLogitsProcessor:
def __init__(self, regex_string: str, tokenizer):
"""Compile the FSM that drives the regex-structured generation.
def adapt_tokenizer(self, tokenizer: PreTrainedTokenizerBase):
"""Adapt vLLM's tokenizer to use to compile the FSM.
Parameters
----------
regex_string
A string that represents a regular expression
tokenizer
The model's tokenizer
The API of Outlines tokenizers is slightly different to that of
`transformers`. The decoder of outlines, returns a list whereas
the decode of vLLM returns an str. To sync the vLLM decoder with
outlines internal api, the decoder should be adapted. In addition
we need to handle the missing spaces to Llama's tokenizer to be
able to compile FSMs for this model.
"""
tokenizer = self.adapt_tokenizer(tokenizer)
fsm = RegexFSM(regex_string, tokenizer)
self.fsm = fsm
if getattr(tokenizer, "_outlines_adapted", False):
return tokenizer
tokenizer.vocabulary = tokenizer.get_vocab()
tokenizer.special_tokens = set(tokenizer.all_special_tokens)
def convert_token_to_string(token: str) -> str:
from transformers.file_utils import SPIECE_UNDERLINE
string = tokenizer.convert_tokens_to_string([token])
# A hack to handle missing spaces to HF's Llama tokenizers
if token.startswith(SPIECE_UNDERLINE) or token == "<0x20>":
return " " + string
return string
def change_decoder(
decoder: Callable[[List[int]], str]
) -> Callable[[List[int]], List[str]]:
"""Sync vLLM's decoder with the outlines by returning list."""
def new_decoder(inp_tokens: List[int]) -> List[str]:
return [decoder(inp_tokens)]
return new_decoder
tokenizer.convert_token_to_string = convert_token_to_string
tokenizer.decode = change_decoder(tokenizer.decode)
setattr(tokenizer, "_outlines_adapted", True) # noqa: B010
return tokenizer
def init_state(self):
"""Initialize the FSM states."""
......@@ -69,50 +99,45 @@ class RegexLogitsProcessor:
return scores
def adapt_tokenizer(self, tokenizer):
"""Adapt vLLM's tokenizer to use to compile the FSM.
The API of Outlines tokenizers is slightly different to that of
`transformers`. In addition we need to handle the missing spaces to
Llama's tokenizer to be able to compile FSMs for this model.
"""
tokenizer.vocabulary = tokenizer.get_vocab()
tokenizer.special_tokens = set(tokenizer.all_special_tokens)
def convert_token_to_string(token: str) -> str:
from transformers.file_utils import SPIECE_UNDERLINE
string = tokenizer.convert_tokens_to_string([token])
class RegexLogitsProcessor(BaseLogitsProcessor):
# A hack to handle missing spaces to HF's Llama tokenizers
if token.startswith(SPIECE_UNDERLINE) or token == "<0x20>":
return " " + string
return string
def __init__(self, regex_string: str, tokenizer: PreTrainedTokenizerBase):
"""Compile the FSM that drives the regex-structured generation.
tokenizer.convert_token_to_string = convert_token_to_string
Parameters
----------
regex_string
A string that represents a regular expression
tokenizer
The model's tokenizer
return tokenizer
"""
tokenizer = self.adapt_tokenizer(tokenizer)
fsm = RegexFSM(regex_string, tokenizer)
self.fsm = fsm
class JSONLogitsProcessor(RegexLogitsProcessor):
def __init__(self,
schema: Union[str, Dict, BaseModel],
tokenizer,
tokenizer: PreTrainedTokenizerBase,
whitespace_pattern: Optional[str] = None):
"""Compile the FSM that drives the JSON-guided generation.
Parameters
----------
schema
A JSON schema that encodes the structure we want the model to generate
A JSON schema that encodes the structure we want the model to
generate
tokenizer
The model's tokenizer
whitespace_pattern
Pattern to use for JSON syntactic whitespace (doesn't impact string literals)
Example: allow only a single space or newline with `whitespace_pattern=r"[\n ]?"`
Pattern to use for JSON syntactic whitespace (doesn't impact
string literals)
Example: allow only a single space or newline with
`whitespace_pattern=r"[\n ]?"`
"""
if isinstance(schema, type(BaseModel)):
schema_str = json.dumps(schema.model_json_schema())
......@@ -122,8 +147,26 @@ class JSONLogitsProcessor(RegexLogitsProcessor):
schema_str = schema
else:
raise ValueError(
f"Cannot parse schema {schema}. The schema must be either " +
"a Pydantic object, a dictionary or a string that contains the JSON "
+ "Schema specification")
f"Cannot parse schema {schema}. The schema must be either "
f"a Pydantic object, a dictionary or a string that contains "
f"the JSON Schema specification")
regex_string = build_regex_from_schema(schema_str, whitespace_pattern)
super().__init__(regex_string, tokenizer)
class CFGLogitsProcessor(BaseLogitsProcessor):
def __init__(self, cfg: str, tokenizer: PreTrainedTokenizerBase):
"""Compile the FSM that drives the context free grammar generation.
Parameters
----------
cfg
A string that represents a context-free grammar
tokenizer
The model's tokenizer
"""
tokenizer = self.adapt_tokenizer(tokenizer)
fsm = CFGFSM(cfg, tokenizer)
self.fsm = fsm
vllm/model_executor/input_metadata.py deleted 100644 → 0
View file @ 2da0dd3e
from typing import Optional
import torch
class InputMetadata:
"""Metadata for input sequences. Used in PagedAttention.
Args:
prompt_lens: Lengths of prompts.
slot_mapping: The address to write the new KV to of each token.
max_context_len: The maximum context length.
context_lens: the length of attention context for each sequence.
block_tables: The block tables. (Seq id -> list of physical block)
kv_cache_dtype: Data type to store kv cache.
"""
def __init__(
self,
is_prompt: bool,
slot_mapping: torch.Tensor,
prompt_lens: Optional[torch.Tensor],
max_seq_len: Optional[int],
start_loc: Optional[torch.Tensor],
max_context_len: Optional[int],
context_lens: Optional[torch.Tensor],
block_tables: Optional[torch.Tensor],
use_cuda_graph: bool,
kv_cache_dtype: str,
) -> None:
self.is_prompt = is_prompt
self.prompt_lens = prompt_lens
self.max_seq_len = max_seq_len
self.start_loc = start_loc
self.max_context_len = max_context_len
self.slot_mapping = slot_mapping
self.context_lens = context_lens
self.block_tables = block_tables
self.use_cuda_graph = use_cuda_graph
self.kv_cache_dtype = kv_cache_dtype
# Set during the execution of the first attention op.
# FIXME(woosuk): This is a hack.
self.attn_bias = None
def __repr__(self) -> str:
return ("InputMetadata("
f"is_prompt={self.is_prompt}, "
f"max_context_len={self.max_context_len}, "
f"slot_mapping={self.slot_mapping}, "
f"context_lens={self.context_lens}, "
f"block_tables={self.block_tables}, "
f"use_cuda_graph={self.use_cuda_graph}, "
f"kv_cache_dtype={self.kv_cache_dtype})")
vllm/model_executor/layers/activation.py
View file @ 7c4f76e3
......@@ -20,8 +20,8 @@ class SiluAndMul(nn.Module):
The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
Shapes:
x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
return: (batch_size, seq_len, d) or (num_tokens, d)
x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
return: (num_tokens, d) or (batch_size, seq_len, d)
"""
def _forward(self, x: torch.Tensor) -> torch.Tensor:
......@@ -47,16 +47,25 @@ class GeluAndMul(nn.Module):
return: (batch_size, seq_len, d) or (num_tokens, d)
"""
def __init__(self, approximate: str = "none"):
super().__init__()
self.approximate = approximate
if approximate not in ("none", "tanh"):
raise ValueError(f"Unknown approximate mode: {approximate}")
def _forward(self, x: torch.Tensor) -> torch.Tensor:
"""PyTorch-native implementation equivalent to forward()."""
d = x.shape[-1] // 2
return F.gelu(x[..., :d]) * x[..., d:]
return F.gelu(x[..., :d], approximate=self.approximate) * x[..., d:]
def forward(self, x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2
output_shape = (x.shape[:-1] + (d, ))
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
ops.gelu_and_mul(out, x)
if self.approximate == "none":
ops.gelu_and_mul(out, x)
elif self.approximate == "tanh":
ops.gelu_tanh_and_mul(out, x)
return out
......
vllm/model_executor/layers/attention.py deleted 100644 → 0
View file @ 2da0dd3e
"""Multi-head attention."""
from typing import List, Optional
import importlib
import torch
import torch.nn as nn
from xformers import ops as xops
from xformers.ops.fmha.attn_bias import (BlockDiagonalCausalMask,
LowerTriangularMaskWithTensorBias)
from vllm._C import ops
from vllm._C import cache_ops
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.triton_kernel.prefix_prefill import (
context_attention_fwd)
from vllm.utils import is_hip
_SUPPORTED_HEAD_SIZES = [64, 80, 96, 112, 128, 256]
# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
_PARTITION_SIZE = 512
class PagedAttention(nn.Module):
"""MHA/MQA/GQA layer with PagedAttention.
This class takes query, key, and value tensors as input. The input tensors
can either contain prompt tokens or generation tokens.
The class does the following:
1. Reshape and store the input key and value tensors in the KV cache.
2. Perform (multi-head/multi-query/grouped-query) attention using either
xformers or the PagedAttention custom op.
3. Return the output tensor.
"""
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
super().__init__()
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = sliding_window
if alibi_slopes is not None:
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.register_buffer("alibi_slopes", alibi_slopes, persistent=False)
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
if self.head_size not in _SUPPORTED_HEAD_SIZES:
raise ValueError(f"head_size ({self.head_size}) is not supported. "
f"Supported head sizes: {_SUPPORTED_HEAD_SIZES}.")
self.use_ref_attention = self.check_use_ref_attention()
def check_use_ref_attention(self) -> bool:
if not is_hip():
return False
# For ROCm, check whether flash attention is installed or not.
# if not, use_ref_attention needs to be True
return importlib.util.find_spec("flash_attn") is None
def ref_masked_attention(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
) -> torch.Tensor:
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
seq_len, _, _ = query.shape
attn_mask = torch.triu(torch.ones(seq_len,
seq_len,
dtype=query.dtype,
device=query.device),
diagonal=1)
attn_mask = attn_mask * torch.finfo(query.dtype).min
attn_weights = self.scale * torch.einsum("qhd,khd->hqk", query,
key).float()
attn_weights = attn_weights + attn_mask.float()
attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
out = torch.einsum("hqk,khd->qhd", attn_weights, value)
return out
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
key_cache: Optional[torch.Tensor],
value_cache: Optional[torch.Tensor],
input_metadata: InputMetadata,
) -> torch.Tensor:
"""PagedAttention forward pass.
Args:
query: shape = [batch_size, seq_len, num_heads * head_size]
key: shape = [batch_size, seq_len, num_kv_heads * head_size]
value: shape = [batch_size, seq_len, num_kv_heads * head_size]
key_cache: shape = [num_blocks, num_kv_heads, head_size/x,
block_size, x]
value_cache: shape = [num_blocks, num_kv_heads, head_size,
block_size]
input_metadata: metadata for the inputs.
Returns:
shape = [batch_size, seq_len, num_heads * head_size]
"""
batch_size, seq_len, hidden_size = query.shape
# Reshape the query, key, and value tensors.
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
# Reshape the keys and values and store them in the cache.
# If key_cache and value_cache are not provided, the new key and value
# vectors will not be cached. This happens during the initial memory
# profiling run.
if key_cache is not None and value_cache is not None:
cache_ops.reshape_and_cache(
key,
value,
key_cache,
value_cache,
input_metadata.slot_mapping.flatten(),
input_metadata.kv_cache_dtype,
)
if input_metadata.is_prompt:
# normal attention
if (key_cache is None or value_cache is None
or input_metadata.block_tables.numel() == 0):
if self.num_kv_heads != self.num_heads:
# As of Nov 2023, xformers only supports MHA. For MQA/GQA,
# project the key and value tensors to the desired number of
# heads.
# TODO(woosuk): Use MQA/GQA kernels for higher performance.
query = query.view(query.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
query.shape[-1])
key = key[:, :,
None, :].expand(key.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
key.shape[-1])
value = value[:, :,
None, :].expand(value.shape[0],
self.num_kv_heads,
self.num_queries_per_kv,
value.shape[-1])
# Set attention bias if not provided. This typically happens at
# the very attention layer of every iteration.
# FIXME(woosuk): This is a hack.
if input_metadata.attn_bias is None:
if self.alibi_slopes is None:
attn_bias = BlockDiagonalCausalMask.from_seqlens(
[seq_len] * batch_size)
if self.sliding_window is not None:
attn_bias = attn_bias.make_local_attention(
self.sliding_window)
input_metadata.attn_bias = attn_bias
else:
input_metadata.attn_bias = _make_alibi_bias(
self.alibi_slopes, self.num_kv_heads, batch_size,
seq_len, query.dtype)
if self.use_ref_attention:
output = self.ref_masked_attention(
query,
key,
value,
)
# Using view got RuntimeError: view size is not compatible with input tensor's size and stride
# (at least one dimension spans across two contiguous subspaces). Use reshape instead
return output.reshape(batch_size, seq_len, hidden_size)
# TODO(woosuk): Too many view operations. Let's try to reduce
# them in the future for code readability.
if self.alibi_slopes is None:
query = query.unsqueeze(0)
key = key.unsqueeze(0)
value = value.unsqueeze(0)
else:
query = query.unflatten(0, (batch_size, seq_len))
key = key.unflatten(0, (batch_size, seq_len))
value = value.unflatten(0, (batch_size, seq_len))
out = xops.memory_efficient_attention_forward(
query,
key,
value,
attn_bias=input_metadata.attn_bias,
p=0.0,
scale=self.scale,
op=xops.fmha.MemoryEfficientAttentionFlashAttentionOp[0] if
(is_hip()) else None,
)
output = out.view_as(query)
else:
# prefix-enabled attention
output = torch.empty_like(query)
context_attention_fwd(
query,
key,
value,
output,
key_cache,
value_cache,
input_metadata.block_tables, # [BS, max_block_per_request]
input_metadata.start_loc,
input_metadata.prompt_lens,
input_metadata.context_lens,
input_metadata.max_seq_len,
getattr(self, "alibi_slopes", None),
)
else:
# Decoding run.
output = _paged_attention(
query,
key_cache,
value_cache,
input_metadata,
self.num_kv_heads,
self.scale,
self.alibi_slopes,
)
# Reshape the output tensor.
return output.view(batch_size, seq_len, hidden_size)
def _make_alibi_bias(
alibi_slopes: torch.Tensor,
num_kv_heads: int,
batch_size: int,
seq_len: int,
dtype: torch.dtype,
) -> LowerTriangularMaskWithTensorBias:
bias = torch.arange(seq_len, dtype=dtype)
# NOTE(zhuohan): HF uses
# `bias = bias[None, :].repeat(prompt_len, 1)`
# here. We find that both biases give the same results, but
# the bias below more accurately follows the original ALiBi
# paper.
bias = bias[None, :] - bias[:, None]
# When using custom attention bias, xformers requires the bias to
# be sliced from a tensor whose length is a multiple of 8.
padded_len = (seq_len + 7) // 8 * 8
num_heads = alibi_slopes.shape[0]
bias = torch.empty(
batch_size,
num_heads,
seq_len,
padded_len,
device=alibi_slopes.device,
dtype=dtype,
)[:, :, :, :seq_len].copy_(bias)
bias.mul_(alibi_slopes[:, None, None])
if num_heads != num_kv_heads:
bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
attn_bias = LowerTriangularMaskWithTensorBias(bias)
return attn_bias
def _paged_attention(
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
input_metadata: InputMetadata,
num_kv_heads: int,
scale: float,
alibi_slopes: Optional[torch.Tensor],
) -> torch.Tensor:
output = torch.empty_like(query)
block_size = value_cache.shape[3]
num_seqs, num_heads, head_size = query.shape
max_num_partitions = (
(input_metadata.max_context_len + _PARTITION_SIZE - 1) //
_PARTITION_SIZE)
# NOTE(woosuk): We use a simple heuristic to decide whether to use
# PagedAttention V1 or V2. If the number of partitions is 1, we use
# V1 to avoid the overhead of reduction. Also, if the number of
# sequences or heads is large, we use V1 since there is enough work
# to parallelize.
# TODO(woosuk): Tune this heuristic.
# For context len > 8192, use V2 kernel to avoid shared memory shortage.
use_v1 = input_metadata.max_context_len <= 8192 and (
max_num_partitions == 1 or num_seqs * num_heads > 512)
if use_v1:
# Run PagedAttention V1.
ops.paged_attention_v1(
output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
input_metadata.block_tables,
input_metadata.context_lens,
block_size,
input_metadata.max_context_len,
alibi_slopes,
input_metadata.kv_cache_dtype,
)
else:
# Run PagedAttention V2.
assert _PARTITION_SIZE % block_size == 0
tmp_output = torch.empty(
size=(num_seqs, num_heads, max_num_partitions, head_size),
dtype=output.dtype,
device=output.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=output.device,
)
max_logits = torch.empty_like(exp_sums)
ops.paged_attention_v2(
output,
exp_sums,
max_logits,
tmp_output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
input_metadata.block_tables,
input_metadata.context_lens,
block_size,
input_metadata.max_context_len,
alibi_slopes,
input_metadata.kv_cache_dtype,
)
return output
vllm/model_executor/layers/fused_moe/__init__.py
View file @ 7c4f76e3
from vllm.model_executor.layers.fused_moe.fused_moe import fused_moe
from vllm.model_executor.layers.fused_moe.fused_moe import (
fused_moe, get_config_file_name)
__all__ = [
"fused_moe",
"get_config_file_name",
]
vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json 0 → 100644
View file @ 7c4f76e3
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json
View file @ 7c4f76e3
{
"1": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 4},
"2": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 7},
"4": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 6},
"8": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 7},
"16": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 7},
"24": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 4},
"32": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 4},
"64": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 4},
"96": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 16, "num_warps": 4, "num_stages": 4},
"128": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 6},
"192": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 6},
"256": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 16, "num_warps": 8, "num_stages": 4},
"512": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 32, "num_warps": 8, "num_stages": 4},
"1024": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 64, "num_warps": 8, "num_stages": 4},
"1536": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 16, "num_warps": 8, "num_stages": 4},
"2048": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 16, "num_warps": 4, "num_stages": 4},
"3072": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 4},
"4096": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 16, "num_warps": 8, "num_stages": 4}
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
}
}
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