Skip to content

GitLab

Commit 09e372e7 authored by zhuwenwen's avatar zhuwenwen
Browse files

add qwen3 moe configs 

support telechat2 and glm4 nn layout
remove log of request_id
parent f2f1b550
Show whitespace changes
Inline Side-by-side
Showing with 580 additions and 8 deletions
vllm/engine/multiprocessing/engine.py
View file @ 09e372e7
......@@ -311,8 +311,8 @@ class MQLLMEngine:
prompt_adapter_request=request.prompt_adapter_request,
priority=request.priority)
if self.log_requests:
logger.info("Added request %s.", request.request_id)
# if self.log_requests:
# logger.info("Added request %s.", request.request_id)
except Exception as e:
# We do not set self._errored = True here, since the error
......
vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=K100_AI_nn.json 0 → 100644
View file @ 09e372e7
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"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": 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": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"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": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"48": {
"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
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"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
},
"128": {
"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
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"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": 5,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"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": 3,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"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=128,N=384,device_name=K100_AI_nn.json 0 → 100644
View file @ 09e372e7
{
"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": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"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
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"32": {
"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
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"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": 3,
"num_ldmatrixes": 1
},
"2048": {
"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
},
"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=128,N=96,device_name=K100_AI_nn.json 0 → 100644
View file @ 09e372e7
{
"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": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"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
},
"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": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"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": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"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": 4,
"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": 3,
"num_ldmatrixes": 1
},
"2048": {
"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
},
"3072": {
"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
},
"4096": {
"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
}
}
vllm/model_executor/model_loader/utils.py
View file @ 09e372e7
......@@ -90,7 +90,7 @@ def get_model_architecture(
architectures = getattr(model_config.hf_config, "architectures", [])
visions = getattr(model_config.hf_config, "visual", []) or getattr(model_config.hf_config, "vision_config", [])
support_nn_architectures = ['LlamaForCausalLM', 'QWenLMHeadModel', 'Qwen2ForCausalLM', 'Qwen2VLForConditionalGeneration', 'Qwen2_5_VLForConditionalGeneration', 'Qwen2MoeForCausalLM', 'Qwen3ForCausalLM', 'Qwen3MoeForCausalLM',
'ChatGLMModel', 'ChatGLMForConditionalGeneration', 'BaichuanForCausalLM', 'BloomForCausalLM', 'MixtralForCausalLM', 'FalconForCausalLM',
'ChatGLMModel', 'Glm4ForCausalLM', 'ChatGLMForConditionalGeneration', 'BaichuanForCausalLM', 'BloomForCausalLM', 'TeleChat2ForCausalLM', 'MixtralForCausalLM', 'FalconForCausalLM',
'MedusaModel', 'MLPSpeculatorPreTrainedModel', 'DeepseekV2ForCausalLM', 'DeepseekV3ForCausalLM', 'DeepSeekMTPModel']
if any(arch in architectures for arch in support_nn_architectures):
if os.getenv('LLAMA_NN') != '0':
......
vllm/model_executor/models/glm4.py
View file @ 09e372e7
......@@ -22,6 +22,8 @@
# limitations under the License.
"""Inference-only GLM-4-0414 model compatible with HuggingFace weights."""
from typing import Iterable, Optional, Set, Tuple, Union
import os
import re
import torch
from torch import nn
......@@ -46,6 +48,9 @@ from .llama import LlamaMLP as Glm4MLP
from .llama import LlamaModel
from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
from vllm.utils import W8a8GetCacheJSON
from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf
class Glm4Attention(nn.Module):
......@@ -270,6 +275,19 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
self.quant_method = None
if quant_config is not None:
self.quant_method=quant_config.get_name()
self.quant_config=quant_config
self.tritonsingleton= W8a8GetCacheJSON()
self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
# self.use_lm_nn = os.environ.get('LM_NN') == '1'
self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
self.use_fa_pad = os.environ.get('FA_PAD') == '1'
self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
......
vllm/model_executor/models/llama.py
View file @ 09e372e7
......@@ -419,6 +419,7 @@ class LlamaModel(nn.Module):
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if self.use_llama_nn:
current_count = loaded_weight.current_count
total_count = loaded_weight.total_count
if "rotary_emb.inv_freq" in name:
......
vllm/model_executor/models/qwen2.py
View file @ 09e372e7
......@@ -393,6 +393,7 @@ class Qwen2Model(nn.Module):
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if self.use_llama_nn:
current_count = loaded_weight.current_count
total_count = loaded_weight.total_count
if "rotary_emb.inv_freq" in name:
......
vllm/model_executor/models/telechat2.py
View file @ 09e372e7
......@@ -21,6 +21,8 @@
# limitations under the License.
from typing import Iterable, Set, Tuple
import os
import re
import torch
import torch.nn as nn
......@@ -31,6 +33,9 @@ from vllm.model_executor.models.llama import LlamaForCausalLM, LlamaModel
from .llama import LlamaDecoderLayer
from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
is_pp_missing_parameter)
from vllm.utils import W8a8GetCacheJSON
from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf
class TeleChat2Model(LlamaModel):
......@@ -50,6 +55,18 @@ class TeleChat2Model(LlamaModel):
layer.mlp.gate_up_proj.bias = None
layer.mlp.gate_up_proj.skip_bias_add = True
self.quant_method = None
if vllm_config.quant_config is not None:
self.quant_method=vllm_config.quant_config.get_name()
self.quant_config=vllm_config.quant_config
self.tritonsingleton= W8a8GetCacheJSON()
self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
self.use_fa_pad = os.environ.get('FA_PAD') == '1'
self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
......@@ -61,6 +78,8 @@ class TeleChat2Model(LlamaModel):
total_num_heads = self.config.n_head
head_dim = self.config.hidden_size // total_num_heads
for name, loaded_weight in weights:
current_count = loaded_weight.current_count
total_count = loaded_weight.total_count
if "self_attn.key_value" in name:
k_weight = []
v_weight = []
......@@ -104,6 +123,47 @@ class TeleChat2Model(LlamaModel):
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
if self.use_llama_nn and self.quant_method is None and current_count==total_count:
lay_key_words = [
"self_attn.qkv_proj.weight",
"self_attn.o_proj.weight",
"mlp.gate_up_proj.weight",
"mlp.down_proj.weight",
]
combined_words = "|".join(lay_key_words)
# lay_qkv_words = ["self_attn.qkv_proj.weight"]
# qkv_words = "|".join(lay_qkv_words)
# for layername, weight in params_dict.items():
# for layername in loaded_params:
for layername in params_dict.keys():
weight = params_dict[layername]
if "lm_head.weight" in layername and weight.shape[1] >= 4096:
lay_key_words.append("lm_head.weight")
combined_words = "|".join(lay_key_words)
os.environ['LM_NN'] = '1'
else:
os.environ['LM_NN'] = '0'
matches = re.findall(combined_words, layername)
if matches:
# if self.use_gemm_pad and gemm_bank_conf(weight.data.shape[0]):
# weight.data = pad_weight(weight.data, 32)
# if self.use_fa_pad and (re.findall(qkv_words, layername)):
# if not gemm_bank_conf(weight.data.shape[0]):
# weight.data = pad_weight(weight.data, 32)
_weight = torch.zeros_like(weight.data)
ori_shape =_weight.shape
ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
weight.data.copy_(_weight)
weight.data=weight.data.reshape(ori_shape[1], -1)
return loaded_params
......
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