Skip to content

GitLab

Commit ce02cd51 authored by dongcl's avatar dongcl
Browse files

Megatron v0.11.0

parent aeed6d97
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1651 additions and 625 deletions
README.md
View file @ ce02cd51
...@@ -66,6 +66,10 @@ def unpermute( ...@@ -66,6 +66,10 @@ def unpermute(
): ):
``` ```
### 项目支持使用[flux kernel](http://10.6.10.68/dcutoolkit/deeplearing/flux)
在tp场景下，用户可以选择使用flux通算融合算子，获得更好的训练和推理性能。项目通过替换transformer engine方法集成flux，使用时需要设置环境变量USE_FLUX_OVERLAP=1，并设置transformer-impl为transformer_engine。
### 使用方式 ### 使用方式
在使用时，进入到examples目录下，有相关模型执行脚本，所用数据集请自行下载：https://r0ddbu55vzx.feishu.cn/drive/folder/ZxHHfCoX4lg75td2hTqcmiAin3g 在使用时，进入到examples目录下，有相关模型执行脚本，所用数据集请自行下载：https://r0ddbu55vzx.feishu.cn/drive/folder/ZxHHfCoX4lg75td2hTqcmiAin3g
``` ```
......
dcu_megatron/adaptor/megatron_adaptor.py
View file @ ce02cd51
...@@ -99,7 +99,7 @@ class CoreAdaptation(MegatronAdaptationABC): ...@@ -99,7 +99,7 @@ class CoreAdaptation(MegatronAdaptationABC):
) )
from ..core.models.gpt.gpt_model import ( from ..core.models.gpt.gpt_model import (
gpt_model_forward, gpt_model_forward,
gpt_model_init, gpt_model_init_wrapper,
shared_embedding_or_mtp_embedding_weight shared_embedding_or_mtp_embedding_weight
) )
from ..training.utils import get_batch_on_this_tp_rank from ..training.utils import get_batch_on_this_tp_rank
...@@ -116,20 +116,20 @@ class CoreAdaptation(MegatronAdaptationABC): ...@@ -116,20 +116,20 @@ class CoreAdaptation(MegatronAdaptationABC):
# GPT Model # GPT Model
MegatronAdaptation.register('megatron.core.models.gpt.gpt_model.GPTModel.forward', gpt_model_forward) MegatronAdaptation.register('megatron.core.models.gpt.gpt_model.GPTModel.forward', gpt_model_forward)
MegatronAdaptation.register('megatron.core.models.gpt.gpt_model.GPTModel.__init__', gpt_model_init) MegatronAdaptation.register('megatron.core.models.gpt.gpt_model.GPTModel.__init__',
gpt_model_init_wrapper,
apply_wrapper=True)
from megatron.core.models.gpt.gpt_model import GPTModel from megatron.core.models.gpt.gpt_model import GPTModel
setattr(GPTModel, 'shared_embedding_or_mtp_embedding_weight', shared_embedding_or_mtp_embedding_weight) setattr(GPTModel, 'shared_embedding_or_mtp_embedding_weight', shared_embedding_or_mtp_embedding_weight)
def patch_core_transformers(self): def patch_core_transformers(self):
from ..core import transformer_block_init_wrapper, transformer_block_forward from ..core import transformer_block_init_wrapper
from ..core.transformer.transformer_config import TransformerConfigPatch, MLATransformerConfigPatch from ..core.transformer.transformer_config import TransformerConfigPatch, MLATransformerConfigPatch
# Transformer block # Transformer block
MegatronAdaptation.register('megatron.core.transformer.transformer_block.TransformerBlock.__init__', MegatronAdaptation.register('megatron.core.transformer.transformer_block.TransformerBlock.__init__',
transformer_block_init_wrapper) transformer_block_init_wrapper)
MegatronAdaptation.register('megatron.core.transformer.transformer_block.TransformerBlock.forward',
transformer_block_forward)
# Transformer config # Transformer config
MegatronAdaptation.register('megatron.core.transformer.transformer_config.TransformerConfig', MegatronAdaptation.register('megatron.core.transformer.transformer_config.TransformerConfig',
...@@ -141,9 +141,9 @@ class CoreAdaptation(MegatronAdaptationABC): ...@@ -141,9 +141,9 @@ class CoreAdaptation(MegatronAdaptationABC):
MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.topk_softmax_with_capacity', MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.topk_softmax_with_capacity',
torch.compile(options={"triton.cudagraphs": True, "triton.cudagraph_trees": False}), torch.compile(options={"triton.cudagraphs": True, "triton.cudagraph_trees": False}),
apply_wrapper=True) apply_wrapper=True)
MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.switch_load_balancing_loss_func', # MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.switch_load_balancing_loss_func',
torch.compile(options={"triton.cudagraphs": True, "triton.cudagraph_trees": False, "triton.cudagraph_support_input_mutation":True}), # torch.compile(options={"triton.cudagraphs": True, "triton.cudagraph_trees": False, "triton.cudagraph_support_input_mutation":True}),
apply_wrapper=True) # apply_wrapper=True)
MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.permute', MegatronAdaptation.register('megatron.core.transformer.moe.moe_utils.permute',
torch.compile(mode='max-autotune-no-cudagraphs'), torch.compile(mode='max-autotune-no-cudagraphs'),
apply_wrapper=True) apply_wrapper=True)
...@@ -166,7 +166,6 @@ class CoreAdaptation(MegatronAdaptationABC): ...@@ -166,7 +166,6 @@ class CoreAdaptation(MegatronAdaptationABC):
def patch_tensor_parallel(self): def patch_tensor_parallel(self):
from ..core.tensor_parallel.cross_entropy import VocabParallelCrossEntropy from ..core.tensor_parallel.cross_entropy import VocabParallelCrossEntropy
from ..core.tensor_parallel import vocab_parallel_embedding_forward, vocab_parallel_embedding_init from ..core.tensor_parallel import vocab_parallel_embedding_forward, vocab_parallel_embedding_init
from ..core.tensor_parallel import ColumnParallelLinearPatch, RowParallelLinearPatch, parallel_linear_init_wrapper
# VocabParallelEmbedding # VocabParallelEmbedding
MegatronAdaptation.register('megatron.core.tensor_parallel.layers.VocabParallelEmbedding.forward', MegatronAdaptation.register('megatron.core.tensor_parallel.layers.VocabParallelEmbedding.forward',
...@@ -188,17 +187,19 @@ class CoreAdaptation(MegatronAdaptationABC): ...@@ -188,17 +187,19 @@ class CoreAdaptation(MegatronAdaptationABC):
apply_wrapper=True) apply_wrapper=True)
# flux # flux
MegatronAdaptation.register("megatron.core.tensor_parallel.layers.ColumnParallelLinear.__init__", if int(os.getenv("USE_FLUX_OVERLAP", "0")):
parallel_linear_init_wrapper, from ..core.tensor_parallel import (
apply_wrapper=True) FluxColumnParallelLinear,
MegatronAdaptation.register("megatron.core.tensor_parallel.layers.ColumnParallelLinear.forward", FluxRowParallelLinear
ColumnParallelLinearPatch.forward) )
MegatronAdaptation.register("megatron.core.tensor_parallel.layers.RowParallelLinear.__init__", from ..core.models.gpt.gpt_layer_specs import get_gpt_layer_with_flux_spec
parallel_linear_init_wrapper,
apply_wrapper=True)
MegatronAdaptation.register("megatron.core.tensor_parallel.layers.RowParallelLinear.forward",
RowParallelLinearPatch.forward)
MegatronAdaptation.register("megatron.core.extensions.transformer_engine.TEColumnParallelLinear",
FluxColumnParallelLinear)
MegatronAdaptation.register("megatron.core.extensions.transformer_engine.TERowParallelLinear",
FluxRowParallelLinear)
MegatronAdaptation.register("megatron.core.models.gpt.gpt_layer_specs.get_gpt_layer_with_transformer_engine_spec",
get_gpt_layer_with_flux_spec)
def patch_training(self): def patch_training(self):
from ..training.tokenizer import build_tokenizer from ..training.tokenizer import build_tokenizer
...@@ -232,19 +233,22 @@ class LegacyAdaptation(MegatronAdaptationABC): ...@@ -232,19 +233,22 @@ class LegacyAdaptation(MegatronAdaptationABC):
self.patch_legacy_models() self.patch_legacy_models()
def patch_legacy_models(self): def patch_legacy_models(self):
from ..legacy.model.transformer import ParallelMLP, ParallelAttention from ..legacy.model.transformer import ParallelMLPPatch, ParallelAttentionPatch
from ..legacy.model.utils import get_norm
# ParallecMLP # ParallecMLP
MegatronAdaptation.register('megatron.legacy.model.transformer.ParallelMLP.__init__', MegatronAdaptation.register('megatron.legacy.model.transformer.ParallelMLP.__init__',
ParallelMLP.__init__) ParallelMLPPatch.__init__)
MegatronAdaptation.register('megatron.legacy.model.transformer.ParallelAttention.forward', MegatronAdaptation.register('megatron.legacy.model.transformer.ParallelAttention.forward',
ParallelAttention.forward) ParallelAttentionPatch.forward)
# rms_norm.RMSNorm # rms_norm.RMSNorm
MegatronAdaptation.register('megatron.legacy.model.rms_norm.RMSNorm.forward', MegatronAdaptation.register('megatron.legacy.model.rms_norm.RMSNorm.forward',
torch.compile(mode="max-autotune-no-cudagraphs"), torch.compile(mode="max-autotune-no-cudagraphs"),
apply_wrapper=True) apply_wrapper=True)
MegatronAdaptation.register('megatron.legacy.model.utils.get_norm',
get_norm)
MegatronAdaptation.execute() MegatronAdaptation.execute()
dcu_megatron/core/__init__.py
View file @ ce02cd51
from .transformer.transformer_block import transformer_block_init_wrapper, transformer_block_forward from .transformer.transformer_block import transformer_block_init_wrapper
dcu_megatron/core/distributed/finalize_model_grads.py
View file @ ce02cd51
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
# Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved.
from typing import List from typing import List
import torch import torch
......
dcu_megatron/core/models/common/embeddings/language_model_embedding.py
View file @ ce02cd51
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
# Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved.
from typing import Literal from typing import Literal
import torch import torch
......
dcu_megatron/core/models/gpt/gpt_layer_specs.py 0 → 100644
View file @ ce02cd51
import warnings
from typing import Optional
from megatron.core.fusions.fused_bias_dropout import get_bias_dropout_add
from megatron.core.models.gpt.moe_module_specs import get_moe_module_spec
from megatron.core.transformer.attention import SelfAttention, SelfAttentionSubmodules
from megatron.core.transformer.enums import AttnMaskType
from megatron.core.transformer.identity_op import IdentityOp
from megatron.core.transformer.mlp import MLP, MLPSubmodules
from megatron.core.transformer.multi_latent_attention import (
MLASelfAttention,
MLASelfAttentionSubmodules,
)
from megatron.core.transformer.spec_utils import ModuleSpec
from megatron.core.transformer.transformer_layer import (
TransformerLayer,
TransformerLayerSubmodules,
)
from dcu_megatron.core.tensor_parallel.layers import FluxColumnParallelLinear, FluxRowParallelLinear
from megatron.core.utils import is_te_min_version
try:
from megatron.core.extensions.transformer_engine import (
TEDotProductAttention,
TENorm,
)
except ImportError:
warnings.warn('transformer_engine is not installed.')
try:
import apex # pylint: disable=unused-import
from megatron.core.fusions.fused_layer_norm import FusedLayerNorm
except ImportError:
warnings.warn('Apex is not installed.')
def get_gpt_layer_with_flux_spec(
num_experts: Optional[int] = None,
moe_grouped_gemm: Optional[bool] = False,
qk_layernorm: Optional[bool] = False,
multi_latent_attention: Optional[bool] = False,
fp8: Optional[str] = None, # pylint: disable=unused-arguments
moe_use_legacy_grouped_gemm: Optional[bool] = False,
) -> ModuleSpec:
"""Use this spec to use flux modules (required for fp8 training).
Args:
num_experts (int, optional): Number of experts. Defaults to None.
moe_grouped_gemm (bool, optional): To use Grouped GEMM. Defaults to False.
qk_layernorm (bool, optional): To use layernorm for queries/keys. Defaults to False.
fp8 (str, optional): Deprecated. For temporary Nemo compatibility.
moe_use_legacy_grouped_gemm (bool, optional): Force use the legacy GroupedMLP.
Defaults to False.
Returns:
ModuleSpec: Module specification with flux modules
"""
if fp8 is not None:
warnings.warn(
'The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated'
' and will be removed soon. Please update your code accordingly.'
)
mlp = get_mlp_module_flux_spec(
use_te=False,
num_experts=num_experts,
moe_grouped_gemm=moe_grouped_gemm,
moe_use_legacy_grouped_gemm=moe_use_legacy_grouped_gemm,
)
if multi_latent_attention:
return ModuleSpec(
module=TransformerLayer,
submodules=TransformerLayerSubmodules(
input_layernorm=TENorm,
self_attention=ModuleSpec(
module=MLASelfAttention,
params={"attn_mask_type": AttnMaskType.causal},
submodules=MLASelfAttentionSubmodules(
linear_q_proj=FluxColumnParallelLinear,
linear_q_down_proj=FluxColumnParallelLinear,
linear_q_up_proj=FluxColumnParallelLinear,
linear_kv_down_proj=FluxColumnParallelLinear,
linear_kv_up_proj=FluxColumnParallelLinear,
core_attention=TEDotProductAttention,
linear_proj=FluxRowParallelLinear,
q_layernorm=TENorm if qk_layernorm else IdentityOp,
kv_layernorm=TENorm if qk_layernorm else IdentityOp,
),
),
self_attn_bda=get_bias_dropout_add,
pre_mlp_layernorm=TENorm,
mlp=mlp,
mlp_bda=get_bias_dropout_add,
),
)
else:
# TENorm significantly harms convergence when used
# for QKLayerNorm if TE Version < 1.9;
# we instead use the Apex implementation.
qk_norm = TENorm if is_te_min_version("1.9.0") else FusedLayerNorm
return ModuleSpec(
module=TransformerLayer,
submodules=TransformerLayerSubmodules(
input_layernorm=TENorm,
self_attention=ModuleSpec(
module=SelfAttention,
params={"attn_mask_type": AttnMaskType.causal},
submodules=SelfAttentionSubmodules(
linear_qkv=FluxColumnParallelLinear,
core_attention=TEDotProductAttention,
linear_proj=FluxRowParallelLinear,
q_layernorm=qk_norm if qk_layernorm else IdentityOp,
k_layernorm=qk_norm if qk_layernorm else IdentityOp,
),
),
self_attn_bda=get_bias_dropout_add,
pre_mlp_layernorm=TENorm,
mlp=mlp,
mlp_bda=get_bias_dropout_add,
),
)
def get_mlp_module_flux_spec(
use_te: Optional[bool] = True,
num_experts: Optional[int] = None,
moe_grouped_gemm: Optional[bool] = False,
fp8: Optional[str] = None, # pylint: disable=unused-arguments
moe_use_legacy_grouped_gemm: Optional[bool] = False,
) -> ModuleSpec:
"""Helper function to get module spec for MLP/MoE"""
if fp8 is not None:
warnings.warn(
'The fp8 argument in "_get_mlp_module_spec" has been deprecated'
' and will be removed soon. Please update your code accordingly.'
)
if num_experts is None:
# Dense MLP w/ or w/o TE modules.
return ModuleSpec(
module=MLP,
submodules=MLPSubmodules(
linear_fc1=FluxColumnParallelLinear,
linear_fc2=FluxRowParallelLinear,
),
)
else:
# Mixture of experts with modules in megatron core.
return get_moe_module_spec(
use_te=True,
num_experts=num_experts,
moe_grouped_gemm=moe_grouped_gemm,
moe_use_legacy_grouped_gemm=moe_use_legacy_grouped_gemm,
)
dcu_megatron/core/models/gpt/gpt_model.py
View file @ ce02cd51
import os
import logging import logging
from typing import Literal, Optional from typing import Literal, Optional
from functools import wraps from functools import wraps
...@@ -16,110 +17,28 @@ from megatron.core.packed_seq_params import PackedSeqParams ...@@ -16,110 +17,28 @@ from megatron.core.packed_seq_params import PackedSeqParams
from megatron.core.transformer.enums import ModelType from megatron.core.transformer.enums import ModelType
from megatron.core.transformer.spec_utils import ModuleSpec from megatron.core.transformer.spec_utils import ModuleSpec
from megatron.core.transformer.transformer_block import TransformerBlock from megatron.core.transformer.transformer_block import TransformerBlock
from megatron.core.extensions.transformer_engine import TEColumnParallelLinear
from dcu_megatron.core.utils import tensor_slide from dcu_megatron.core.utils import tensor_slide
from dcu_megatron.core.transformer.mtp.multi_token_predictor import MultiTokenPredictor from dcu_megatron.core.transformer.mtp.multi_token_predictor import MultiTokenPredictor
from dcu_megatron.core.transformer.transformer_config import TransformerConfig from dcu_megatron.core.transformer.transformer_config import TransformerConfig
from dcu_megatron.core.tensor_parallel import FluxColumnParallelLinear
def gpt_model_init( def gpt_model_init_wrapper(fn):
self, @wraps(fn)
config: TransformerConfig, def wrapper(self, *args, **kwargs):
transformer_layer_spec: ModuleSpec, fn(self, *args, **kwargs)
vocab_size: int,
max_sequence_length: int,
pre_process: bool = True,
post_process: bool = True,
fp16_lm_cross_entropy: bool = False,
parallel_output: bool = True,
share_embeddings_and_output_weights: bool = False,
position_embedding_type: Literal['learned_absolute', 'rope', 'none'] = 'learned_absolute',
rotary_percent: float = 1.0,
rotary_base: int = 10000,
rope_scaling: bool = False,
rope_scaling_factor: float = 8.0,
scatter_embedding_sequence_parallel: bool = True,
seq_len_interpolation_factor: Optional[float] = None,
mtp_spec: ModuleSpec = None
) -> None:
super(GPTModel, self).__init__(config=config)
if has_config_logger_enabled(config):
log_config_to_disk(config, locals(), prefix=type(self).__name__)
self.transformer_layer_spec: ModuleSpec = transformer_layer_spec
self.vocab_size = vocab_size
self.max_sequence_length = max_sequence_length
self.pre_process = pre_process
self.post_process = post_process
self.fp16_lm_cross_entropy = fp16_lm_cross_entropy
self.parallel_output = parallel_output
self.share_embeddings_and_output_weights = share_embeddings_and_output_weights
self.position_embedding_type = position_embedding_type
# megatron core pipelining currently depends on model type
# TODO: remove this dependency ?
self.model_type = ModelType.encoder_or_decoder
# These 4 attributes are needed for TensorRT-LLM export.
self.max_position_embeddings = max_sequence_length
self.rotary_percent = rotary_percent
self.rotary_base = rotary_base
self.rotary_scaling = rope_scaling
if self.pre_process:
self.embedding = LanguageModelEmbedding(
config=self.config,
vocab_size=self.vocab_size,
max_sequence_length=self.max_sequence_length,
position_embedding_type=position_embedding_type,
scatter_to_sequence_parallel=scatter_embedding_sequence_parallel,
)
if self.position_embedding_type == 'rope' and not self.config.multi_latent_attention:
self.rotary_pos_emb = RotaryEmbedding(
kv_channels=self.config.kv_channels,
rotary_percent=rotary_percent,
rotary_interleaved=self.config.rotary_interleaved,
seq_len_interpolation_factor=seq_len_interpolation_factor,
rotary_base=rotary_base,
rope_scaling=rope_scaling,
rope_scaling_factor=rope_scaling_factor,
use_cpu_initialization=self.config.use_cpu_initialization,
)
# Cache for RoPE tensors which do not change between iterations.
self.rotary_pos_emb_cache = {}
# Transformer.
self.decoder = TransformerBlock(
config=self.config,
spec=transformer_layer_spec,
pre_process=self.pre_process,
post_process=self.post_process
)
# Output
if post_process:
if self.config.defer_embedding_wgrad_compute:
# The embedding activation buffer preserves a reference to the input activations
# of the final embedding projection layer GEMM. It will hold the activations for
# all the micro-batches of a global batch for the last pipeline stage. Once we are
# done with all the back props for all the microbatches for the last pipeline stage,
# it will be in the pipeline flush stage. During this pipeline flush we use the
# input activations stored in embedding activation buffer and gradient outputs
# stored in gradient buffer to calculate the weight gradients for the embedding
# final linear layer.
self.embedding_activation_buffer = []
self.grad_output_buffer = []
else:
self.embedding_activation_buffer = None
self.grad_output_buffer = None
self.output_layer = tensor_parallel.ColumnParallelLinear( if (
config.hidden_size, self.post_process
and int(os.getenv("USE_FLUX_OVERLAP", "0"))
):
self.output_layer = FluxColumnParallelLinear(
self.config.hidden_size,
self.vocab_size, self.vocab_size,
config=config, config=self.config,
init_method=config.init_method, init_method=self.config.init_method,
bias=False, bias=False,
skip_bias_add=False, skip_bias_add=False,
gather_output=not self.parallel_output, gather_output=not self.parallel_output,
...@@ -129,18 +48,22 @@ def gpt_model_init( ...@@ -129,18 +48,22 @@ def gpt_model_init(
grad_output_buffer=self.grad_output_buffer, grad_output_buffer=self.grad_output_buffer,
) )
self.setup_embeddings_and_output_layer()
# add mtp # add mtp
self.mtp_spec: ModuleSpec = mtp_spec
self.num_nextn_predict_layers = self.config.num_nextn_predict_layers self.num_nextn_predict_layers = self.config.num_nextn_predict_layers
if self.num_nextn_predict_layers:
assert hasattr(self.config, "mtp_spec")
self.mtp_spec: ModuleSpec = self.config.mtp_spec
self.share_mtp_embedding_and_output_weight = self.config.share_mtp_embedding_and_output_weight self.share_mtp_embedding_and_output_weight = self.config.share_mtp_embedding_and_output_weight
self.recompute_mtp_norm = self.config.recompute_mtp_norm self.recompute_mtp_norm = self.config.recompute_mtp_norm
self.recompute_mtp_layer = self.config.recompute_mtp_layer self.recompute_mtp_layer = self.config.recompute_mtp_layer
self.mtp_loss_scale = self.config.mtp_loss_scale self.mtp_loss_scale = self.config.mtp_loss_scale
if self.post_process and self.training and self.num_nextn_predict_layers: if self.post_process and self.training:
self.mtp_layers = torch.nn.ModuleList( self.mtp_layers = torch.nn.ModuleList(
[ [
MultiTokenPredictor( MultiTokenPredictor(
config, self.config,
self.mtp_spec.submodules, self.mtp_spec.submodules,
vocab_size=self.vocab_size, vocab_size=self.vocab_size,
max_sequence_length=self.max_sequence_length, max_sequence_length=self.max_sequence_length,
...@@ -161,16 +84,10 @@ def gpt_model_init( ...@@ -161,16 +84,10 @@ def gpt_model_init(
) )
if self.pre_process or self.post_process: if self.pre_process or self.post_process:
self.setup_embeddings_and_output_layer()
if has_config_logger_enabled(self.config):
log_config_to_disk(
self.config, self.state_dict(), prefix=f'{type(self).__name__}_init_ckpt'
)
if self.num_nextn_predict_layers and (self.pre_process or self.post_process):
setup_mtp_embeddings(self) setup_mtp_embeddings(self)
return wrapper
def shared_embedding_or_mtp_embedding_weight(self) -> Tensor: def shared_embedding_or_mtp_embedding_weight(self) -> Tensor:
"""Gets the embedding weight when share embedding and mtp embedding weights set to True. """Gets the embedding weight when share embedding and mtp embedding weights set to True.
...@@ -424,10 +341,10 @@ def gpt_model_forward( ...@@ -424,10 +341,10 @@ def gpt_model_forward(
if ( if (
self.num_nextn_predict_layers self.num_nextn_predict_layers
and getattr(self.decoder, "final_layernorm", None) is not None and getattr(self.decoder, "main_final_layernorm", None) is not None
): ):
# move block main model final norms here # move block main model final norms here
hidden_states = self.decoder.final_layernorm(hidden_states) hidden_states = self.decoder.main_final_layernorm(hidden_states)
logits, _ = self.output_layer( logits, _ = self.output_layer(
hidden_states, weight=output_weight, runtime_gather_output=runtime_gather_output hidden_states, weight=output_weight, runtime_gather_output=runtime_gather_output
......
dcu_megatron/core/tensor_parallel/__init__.py
View file @ ce02cd51
from .layers import ( from .layers import (
parallel_linear_init_wrapper, FluxColumnParallelLinear,
ColumnParallelLinearPatch, FluxRowParallelLinear,
RowParallelLinearPatch,
vocab_parallel_embedding_forward, vocab_parallel_embedding_forward,
vocab_parallel_embedding_init, vocab_parallel_embedding_init,
) )
\ No newline at end of file
dcu_megatron/core/tensor_parallel/layers.py
View file @ ce02cd51
This diff is collapsed.
dcu_megatron/core/transformer/mtp/multi_token_predictor.py
View file @ ce02cd51
# Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved. import os
import logging import logging
from dataclasses import dataclass from dataclasses import dataclass
from typing import Union, Optional, Literal from typing import Union, Optional, Literal
...@@ -11,6 +11,7 @@ from megatron.core.models.common.embeddings.language_model_embedding import Lang ...@@ -11,6 +11,7 @@ from megatron.core.models.common.embeddings.language_model_embedding import Lang
from megatron.core.models.common.embeddings.rotary_pos_embedding import RotaryEmbedding from megatron.core.models.common.embeddings.rotary_pos_embedding import RotaryEmbedding
from megatron.core.packed_seq_params import PackedSeqParams from megatron.core.packed_seq_params import PackedSeqParams
from megatron.core.transformer.module import MegatronModule from megatron.core.transformer.module import MegatronModule
from megatron.core.extensions.transformer_engine import TEColumnParallelLinear
from megatron.core.fusions.fused_cross_entropy import fused_vocab_parallel_cross_entropy from megatron.core.fusions.fused_cross_entropy import fused_vocab_parallel_cross_entropy
from megatron.core.transformer import ModuleSpec, TransformerConfig, build_module from megatron.core.transformer import ModuleSpec, TransformerConfig, build_module
...@@ -136,12 +137,16 @@ class MultiTokenPredictor(MegatronModule): ...@@ -136,12 +137,16 @@ class MultiTokenPredictor(MegatronModule):
self.embedding_activation_buffer = None self.embedding_activation_buffer = None
self.grad_output_buffer = None self.grad_output_buffer = None
self.output_layer = tensor_parallel.ColumnParallelLinear( if int(os.getenv("USE_FLUX_OVERLAP", "0")):
config.hidden_size, column_parallel_linear_impl = FluxColumnParallelLinear
else:
column_parallel_linear_impl = tensor_parallel.ColumnParallelLinear
self.output_layer = column_parallel_linear_impl(
self.config.hidden_size,
self.vocab_size, self.vocab_size,
config=config, config=self.config,
init_method=config.init_method, init_method=self.config.init_method,
bias=self.add_output_layer_bias, bias=False,
skip_bias_add=False, skip_bias_add=False,
gather_output=not self.parallel_output, gather_output=not self.parallel_output,
skip_weight_param_allocation=self.share_mtp_embedding_and_output_weight, skip_weight_param_allocation=self.share_mtp_embedding_and_output_weight,
......
dcu_megatron/core/transformer/transformer_block.py
View file @ ce02cd51
from contextlib import nullcontext
from typing import Optional
from functools import wraps from functools import wraps
import torch
from torch import Tensor
from megatron.core import InferenceParams, parallel_state, tensor_parallel
from megatron.core.fusions.fused_layer_norm import FusedLayerNorm
from megatron.core.packed_seq_params import PackedSeqParams
from megatron.core.utils import make_viewless_tensor
try:
from megatron.core.extensions.transformer_engine import TEDelayedScaling
HAVE_TE = True
except ImportError:
HAVE_TE = False
def transformer_block_init_wrapper(fn): def transformer_block_init_wrapper(fn):
@wraps(fn) @wraps(fn)
...@@ -25,178 +8,8 @@ def transformer_block_init_wrapper(fn): ...@@ -25,178 +8,8 @@ def transformer_block_init_wrapper(fn):
# mtp require seperate layernorms for main model and mtp modules, thus move finalnorm out of block # mtp require seperate layernorms for main model and mtp modules, thus move finalnorm out of block
config = args[0] if len(args) > 1 else kwargs['config'] config = args[0] if len(args) > 1 else kwargs['config']
self.move_final_norm_out_of_block = getattr(config, "num_nextn_predict_layers", 0) > 0 if getattr(config, "num_nextn_predict_layers", 0) > 0:
self.main_final_layernorm = self.final_layernorm
self.final_layernorm = None
return wrapper return wrapper
def transformer_block_forward(
self,
hidden_states: Tensor,
attention_mask: Tensor,
context: Tensor = None,
context_mask: Tensor = None,
rotary_pos_emb: Tensor = None,
rotary_pos_cos: Tensor = None,
rotary_pos_sin: Tensor = None,
attention_bias: Tensor = None,
inference_params: InferenceParams = None,
packed_seq_params: PackedSeqParams = None,
sequence_len_offset: Tensor = None,
):
"""
Perform the forward pass through the transformer block.
This method handles the core computation of the transformer, including
self-attention, optional cross-attention, and feed-forward operations.
Args:
hidden_states (Tensor): Input tensor of shape [s, b, h] where s is the
sequence length, b is the batch size, and h is the hidden size.
attention_mask (Tensor): Boolean tensor of shape [1, 1, s, s] for masking
self-attention.
context (Tensor, optional): Context tensor for cross-attention.
context_mask (Tensor, optional): Mask for cross-attention context
rotary_pos_emb (Tensor, optional): Rotary positional embeddings.
attention_bias (Tensor): Bias tensor for Q * K.T of shape in shape broadcastable
to [b, num_head, sq, skv], e.g. [1, 1, sq, skv].
Used as an alternative to apply attention mask for TE cuDNN attention.
inference_params (InferenceParams, optional): Parameters for inference-time
optimizations.
packed_seq_params (PackedSeqParams, optional): Parameters for packed sequence
processing.
Returns:
Union[Tensor, Tuple[Tensor, Tensor]]: The output hidden states tensor of shape
[s, b, h], and optionally the updated context tensor if cross-attention is used.
"""
if not self.pre_process:
# See set_input_tensor()
hidden_states = self.input_tensor
# Update the inference parameters with the current batch size in case it is variable
if inference_params and not self.training:
inference_params.current_batch_size = hidden_states.size(1)
# Viewless tensor.
# - We only need to create a viewless tensor in the case of micro batch
# size (mbs) == 1, since in this case, 'hidden_states.transpose()'
# above creates a view tensor, and '.contiguous()' is a pass-through.
# For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
# the need to make it viewless.
#
# However, we don't explicitly check mbs == 1 here because
# make_viewless_tensor() has negligible overhead when its input
# is already viewless.
#
# - For the 'else' case above, calling make_viewless_tensor() here is
# likely redundant, since p2p_communication.py (likely originator)
# already creates viewless tensors. That said, make_viewless_tensor()
# is called here to be future-proof and corner-case-proof.
hidden_states = make_viewless_tensor(inp=hidden_states, requires_grad=True, keep_graph=True)
if self.config.sequence_parallel:
rng_context = tensor_parallel.get_cuda_rng_tracker().fork()
else:
rng_context = nullcontext()
if self.config.fp8:
import transformer_engine # To keep out TE dependency when not training in fp8
if self.config.fp8 == "e4m3":
fp8_format = transformer_engine.common.recipe.Format.E4M3
elif self.config.fp8 == "hybrid":
fp8_format = transformer_engine.common.recipe.Format.HYBRID
else:
raise ValueError("E4M3 and HYBRID are the only supported FP8 formats.")
fp8_recipe = TEDelayedScaling(
config=self.config,
fp8_format=fp8_format,
override_linear_precision=(False, False, not self.config.fp8_wgrad),
)
fp8_group = None
if parallel_state.model_parallel_is_initialized():
fp8_group = parallel_state.get_amax_reduction_group(
with_context_parallel=True, tp_only_amax_red=self.tp_only_amax_red
)
fp8_context = transformer_engine.pytorch.fp8_autocast(
enabled=True, fp8_recipe=fp8_recipe, fp8_group=fp8_group
)
else:
fp8_context = nullcontext()
with rng_context, fp8_context:
# Forward pass.
if self.config.recompute_granularity == 'full' and self.training:
hidden_states = self._checkpointed_forward(
hidden_states=hidden_states,
attention_mask=attention_mask,
context=context,
context_mask=context_mask,
rotary_pos_emb=rotary_pos_emb,
attention_bias=attention_bias,
packed_seq_params=packed_seq_params,
)
else:
for l_no, layer in enumerate(self.layers):
with self.offload_context:
layer.use_cudagraph = True
if (len(self.cuda_graphs) == 0) or (not self.training):
hidden_states, context = layer(
hidden_states=hidden_states,
attention_mask=attention_mask,
context=context,
context_mask=context_mask,
rotary_pos_emb=rotary_pos_emb,
rotary_pos_cos=rotary_pos_cos,
rotary_pos_sin=rotary_pos_sin,
attention_bias=attention_bias,
inference_params=inference_params,
packed_seq_params=packed_seq_params,
sequence_len_offset=sequence_len_offset,
)
else:
# CUDA graph replay for layer `l_no` and microbatch
# `self.current_microbatch`. TransformerEngine versions>=1.10
# allow keyword arguments with CUDA graph. However, CUDA graph
# acccepts only Tensor inputs and Tensor outputs. Hence,
# `inference_params` and `packed_seq_params` are excluded from
# input list while output is limited to `hidden_states`.
cg_index = self.current_microbatch % len(self.cuda_graphs[l_no])
assert not any(
[inference_params, packed_seq_params]
), "CUDA graph accepts only Tensor inputs."
optional_inputs = self.get_cuda_graph_optional_args(
attention_mask,
context,
context_mask,
rotary_pos_emb,
attention_bias,
inference_params,
packed_seq_params,
)
hidden_states = self.cuda_graphs[l_no][cg_index](
hidden_states, **optional_inputs
)
if (
torch.is_grad_enabled()
and self.config.cpu_offloading
and self.group_prefetch_offload_commit_async is not None
):
hidden_states = self.group_prefetch_offload_commit_async(hidden_states)
# Final layer norm.
if (not self.move_final_norm_out_of_block) and self.final_layernorm is not None:
hidden_states = self.final_layernorm(hidden_states)
# TENorm produces a "viewed" tensor. This will result in schedule.py's
# deallocate_output_tensor() throwing an error, so a viewless tensor is
# created to prevent this.
hidden_states = make_viewless_tensor(
inp=hidden_states, requires_grad=True, keep_graph=True
)
return hidden_states
dcu_megatron/core/transformer/transformer_config.py
View file @ ce02cd51
...@@ -26,9 +26,6 @@ class ExtraTransformerConfig: ...@@ -26,9 +26,6 @@ class ExtraTransformerConfig:
################## ##################
# flux # flux
################## ##################
use_flux: bool = False
"""If set, flux will be used in ColumnParallelLinear and RowParallelLinear"""
flux_transpose_weight: bool = False flux_transpose_weight: bool = False
......
dcu_megatron/core/utils.py
View file @ ce02cd51
import torch import torch
from typing import List, Optional, Union from typing import List, Optional, Union
from importlib.metadata import version
from packaging.version import Version as PkgVersion
_flux_version = None
def get_flux_version():
"""Get flux version from __version__; if not available use pip's. Use caching."""
def get_flux_version_str():
import flux
if hasattr(flux, '__version__'):
return str(flux.__version__)
else:
return version("flux")
global _flux_version
if _flux_version is None:
_flux_version = PkgVersion(get_flux_version_str())
return _flux_version
def is_flux_min_version(version, check_equality=True):
"""Check if minimum version of `flux` is installed."""
if check_equality:
return get_flux_version() >= PkgVersion(version)
return get_flux_version() > PkgVersion(version)
def tensor_slide( def tensor_slide(
......
dcu_megatron/legacy/model/rms_norm.py
View file @ ce02cd51
# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
import torch
from torch import nn
class RMSNorm(torch.nn.Module):
def __init__(self,
dim: int,
eps: float = 1e-6,
sequence_parallel: bool = False,
config: dict = None):
"""RMS Normaliation module
Args:
dim (int): The width of input, i.e. hidden size
eps (float): epsilon to use for the norm, default to 1e-6
sequence_parallel (bool): Set to true if sequence parallelism is being used,
this marks the weights as needing to be allreduced.
"""
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
setattr(self.weight, 'sequence_parallel', sequence_parallel)
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
@torch.compile(mode="max-autotune-no-cudagraphs")
def forward(self, x):
output = self._norm(x.float()).type_as(x)
return output * self.weight
import torch
from typing import Any, Callable, Dict, Optional, Tuple, Union
import lightop # rmsnorm_forward,rmsnorm_backward
from functools import partial
from megatron.core.utils import is_torch_min_version
if is_torch_min_version("2.4.0a0"):
custom_fwd = partial(torch.amp.custom_fwd, device_type="cuda")
custom_bwd = partial(torch.amp.custom_bwd, device_type="cuda")
else:
custom_fwd = torch.cuda.amp.custom_fwd
custom_bwd = torch.cuda.amp.custom_bwd
class _LightopRMSNorm(torch.autograd.Function):
""" 使用lightop实现rmsnorm"""
@staticmethod
# @custom_fwd
def forward(ctx,
inp: torch.Tensor,
weight: torch.Tensor,
ln_out: torch.Tensor,
eps: float,
is_grad_enabled):
output = lightop.rmsnorm_forward(inp, weight, ln_out, eps, training=True)# output = (output, weight)
rsigma = output[1]
if is_grad_enabled:
ctx.save_for_backward(inp, weight, rsigma)
return output[0]
@staticmethod
# @custom_bwd
def backward(ctx, grad_output):
inp, weight, rsigma = ctx.saved_tensors
dgrad, dgamma = lightop.rmsnorm_backward(grad_output, inp, rsigma, weight)
return dgrad, dgamma, None, None, None
class LightopRMSNorm(torch.nn.Module):
def __init__(self,
dim: int,
eps: float = 1e-6,):
"""RMS Normaliation module
Args:
dim (int): The width of input, i.e. hidden size
eps (float): epsilon to use for the norm, default to 1e-6
"""
super().__init__()
self.eps = eps
self.weight = torch.nn.Parameter(torch.ones(dim))
# @no_torch_dynamo() # 动态torch._dynamo.disable
def forward(self, inp: torch.Tensor, is_first_microbatch: Optional[bool] = None):
if torch.is_grad_enabled():
fwd_fn = _LightopRMSNorm.apply
args = []
else:
fwd_fn = _LightopRMSNorm.forward
args = [None]
ln_out = torch.empty_like(inp, dtype=inp.dtype, memory_format=torch.contiguous_format)
args += (inp, self.weight, ln_out, self.eps, torch.is_grad_enabled())
out = fwd_fn(*args)
return out
dcu_megatron/legacy/model/transformer.py
View file @ ce02cd51
...@@ -3,14 +3,21 @@ import torch.nn.functional as F ...@@ -3,14 +3,21 @@ import torch.nn.functional as F
from megatron.training import get_args from megatron.training import get_args
from megatron.core import tensor_parallel from megatron.core import tensor_parallel
from megatron.legacy.model.enums import AttnType
from megatron.core.models.common.embeddings import apply_rotary_pos_emb
from megatron.legacy.model.module import MegatronModule from megatron.legacy.model.module import MegatronModule
from megatron.legacy.model.transformer import ParallelMLP
from megatron.legacy.model.utils import ( from megatron.legacy.model.utils import (
erf_gelu, erf_gelu,
openai_gelu, openai_gelu,
) )
try:
from einops import rearrange
except ImportError:
rearrange = None
class ParallelMLP(MegatronModule): class ParallelMLPPatch(MegatronModule):
"""MLP. """MLP.
MLP will take the input with h hidden state, project it to 4*h MLP will take the input with h hidden state, project it to 4*h
...@@ -74,7 +81,7 @@ class ParallelMLP(MegatronModule): ...@@ -74,7 +81,7 @@ class ParallelMLP(MegatronModule):
) )
class ParallelAttention(MegatronModule): class ParallelAttentionPatch(MegatronModule):
"""Parallel self-attention layer abstract class. """Parallel self-attention layer abstract class.
Self-attention layer takes input with size [s, b, h] Self-attention layer takes input with size [s, b, h]
......
dcu_megatron/legacy/model/utils.py 0 → 100644
View file @ ce02cd51
from megatron.training import get_args
from megatron.legacy.model import LayerNorm
from .rms_norm import RMSNorm, LightopRMSNorm
def get_norm(config):
args = get_args()
if args.normalization == "LayerNorm":
return LayerNorm(
config.hidden_size,
eps=config.layernorm_epsilon,
no_persist_layer_norm=not config.persist_layer_norm,
sequence_parallel=config.sequence_parallel,
apply_layernorm_1p=args.apply_layernorm_1p)
elif args.normalization == "RMSNorm":
if args.apply_layernorm_1p:
raise NotImplementedError('RMSNorm does not currently support the layernorm_1p formulation.')
return RMSNorm(dim=config.hidden_size,
eps=config.layernorm_epsilon,
sequence_parallel=config.sequence_parallel)
elif args.normalization == "LightopRMSNorm":
return LightopRMSNorm(dim=config.hidden_size,
eps=config.layernorm_epsilon)
else:
raise Exception(f"unsupported norm type '{args.normalization}'.")
dcu_megatron/training/arguments.py
View file @ ce02cd51
...@@ -51,6 +51,7 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False): ...@@ -51,6 +51,7 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
# Standard arguments. # Standard arguments.
parser = _add_network_size_args(parser) parser = _add_network_size_args(parser)
parser = _add_extra_network_size_args(parser)
parser = _add_regularization_args(parser) parser = _add_regularization_args(parser)
parser = _add_training_args(parser) parser = _add_training_args(parser)
parser = _add_extra_training_args(parser) parser = _add_extra_training_args(parser)
...@@ -106,6 +107,18 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False): ...@@ -106,6 +107,18 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
return args return args
def _add_extra_network_size_args(parser):
# 删除原参数
remove_original_params(parser, ["normalization"])
# 重定义参数
group = parser.add_argument_group(title='extra network size args')
group.add_argument('--normalization', default='LayerNorm',
choices=['LayerNorm', 'RMSNorm', 'LightopRMSNorm'],
help='Which normalization technique to use.')
return parser
def _add_extra_distributed_args(parser): def _add_extra_distributed_args(parser):
group = parser.add_argument_group(title='extra distributed args') group = parser.add_argument_group(title='extra distributed args')
group.add_argument('--rank', default=-1, type=int, group.add_argument('--rank', default=-1, type=int,
...@@ -169,9 +182,7 @@ def _add_mtp_args(parser): ...@@ -169,9 +182,7 @@ def _add_mtp_args(parser):
def _add_flux_args(parser): def _add_flux_args(parser):
group = parser.add_argument_group(title='multi token prediction') group = parser.add_argument_group(title='flux args')
group.add_argument('--use-flux', action='store_true', default=False,
help='If set, flux will be used in ColumnParallelLinear and RowParallelLinear')
group.add_argument('--flux-transpose-weight', action='store_true', default=False, group.add_argument('--flux-transpose-weight', action='store_true', default=False,
help='Whether to transpose weight when using flux kernel') help='Whether to transpose weight when using flux kernel')
return parser return parser
examples/llama/Llama2_70b.sh 0 → 100755
View file @ ce02cd51
#!/bin/bash
# set -eux
#export FLASH_ATTENTION_PRINT_PARAM=1
# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB # SYS
#export HIP_ALLOC_INITIALIZE=0
# export GPU_MAX_HW_QUEUES=10
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志
source /opt/dtk/env.sh
# 导入hipblaslt库
# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
# 更新rocblas
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1
# prof采集添加同步, 避免卡顿
# export GPU_FLUSH_ON_EXECUTION=1
# export HIP_DIRECT_DISPATCH=0
# 采集rocblas size
# export ROCBLAS_LAYER=3
# export HIPBLASLT_LOG_LEVEL=3
# 采集 fa size
# export FLASH_ATTENTION_PRINT_PARAM=1
#增加编译缓存
export cache_size_limit=64
# lightop算子库
export PYTORCH_ROCM_ARCH='gfx906;gfx926;gfx936'
# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b #
SAVE_PATH=./tmp_7b
TENSORBOARD_LOGS_PATH=./tmp_7b #$2 #<Specify path>
DATA_PATH="/public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/redpajama_text_document" #<Specify path and file prefix>_text_document
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
GPT_MODEL_ARGS=(
--num-layers 80 #80 #80 #40 # 20 #
--hidden-size 8192
--ffn-hidden-size 22016 # 28672
--num-attention-heads 64
--max-position-embeddings 8192
--group-query-attention
--num-query-groups 8
--normalization RMSNorm
--position-embedding-type rope
--untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
)
export NVTE_FLASH_ATTN=1 # 走cutlass
# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
# --transformer-impl transformer_engine # 走core用这两组参数
# --use-mcore-models
# --transformer-impl local # 走legacy用这两组参数
# --use-legacy-models
TRAINING_ARGS=(
--transformer-impl local # 走legacy用这两组参数
--use-legacy-models
--micro-batch-size 1
--global-batch-size 512 #32 #512 #256 # 64 #240 #60 #512 #64
--train-iters 300
--weight-decay 0.1
--adam-beta1 0.9
--adam-beta2 0.95
--init-method-std 0.006
--clip-grad 1.0
--bf16
# --fp16 # 开启fp16需要指定loss-scale
# --loss-scale 1024
--use-distributed-optimizer
--disable-bias-linear
--attention-dropout 0
--hidden-dropout 0
# --no-gradient-accumulation-fusion
# --no-check-for-nan-in-loss-and-grad
--swiglu
--lr 3.0e-5
--lr-decay-style cosine
--min-lr 3.0e-6
--lr-warmup-iters 1
--ckpt-format torch
--ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
# --recompute-activations
# --recompute-granularity full # 开启重计算降低显存增加耗时
# --recompute-num-layers 1 #0 #
# --recompute-method block
--overlap-grad-reduce # 重叠ddp grad reduce
# --tp-comm-overlap # tensor parallel comm和gemm重叠, 启动core
# --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 启动core
--use-flash-attn
)
# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
# export TORCHINDUCTOR_BENCHMARK_FUSION=1
# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
# export TORCHINDUCTOR_MAX_AUTOTUNE=1
# export TORCHINDUCTOR_CACHE_DIR=./cache
# --use-flash-attn-cutlass # cutlass fa
# --use-flash-attn-triton # triton fa
# --use-flash-attn-torch # torch fa
MODEL_PARALLEL_ARGS=(
--sequence-parallel
--tensor-model-parallel-size 4
--pipeline-model-parallel-size 8
--context-parallel-size 1
# --num-layers-per-virtual-pipeline-stage 1
# --microbatch-group-size-per-virtual-pipeline-stage 5
# --no-overlap-p2p-communication # 开启后
)
DATA_ARGS=(
--data-path $DATA_PATH
--seq-length 4096 #8192 #4096
--split 949,50,1
--tokenizer-type Llama2Tokenizer
--tokenizer-model /public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/tokenizer.model
# --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
)
EVAL_AND_LOGGING_ARGS=(
--log-interval 1
--log-throughput
--save-interval 500
--eval-interval 50
--eval-iters 3
--save $SAVE_PATH
--load $SAVE_PATH
--tensorboard-dir $TENSORBOARD_LOGS_PATH
)
# FINETUNE_ARGS=(
# # --finetune
# # --pretrained-checkpoint $CHECKPOINT_PATH
# --load $CHECKPOINT_PATH
# --no-load-optim
# --no-load-rng
# )
PROFILE_ARGS=(
--profile
--profile-step-start 4
--profile-step-end 5
--use-pytorch-profiler
--profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
--profile-dir prof_data
)
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=34577
DISTRIBUTED_ARGS=(
--rank ${RANK}
--world-size ${WORLD_SIZE}
--local-rank ${LOCAL_RANK}
--dist-url tcp://${DIST_URL}:${DIST_PORT}
)
APP="python -u ../../pretrain_gpt.py \
${GPT_MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${MODEL_PARALLEL_ARGS[@]} \
${DATA_ARGS[@]} \
${EVAL_AND_LOGGING_ARGS[@]} \
${DISTRIBUTED_ARGS[@]} \
"
# 开启profile
# ${PROFILE_ARGS[@]} \
# export HIP_VISIBLE_DEVICES=0,7 # # 4,5,6,7 #,
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # # 4,5,6,7 #,
# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
# ${APP}
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
esac
\ No newline at end of file
examples/llama/Llama2_7b.sh 0 → 100755
View file @ ce02cd51
#!/bin/bash
# set -eux
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
#default env
#export FLASH_ATTENTION_PRINT_PARAM=1
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB # SYS
# export GPU_MAX_HW_QUEUES=10
#export HIP_ALLOC_INITIALIZE=0
export CUDA_DEVICE_MAX_CONNECTIONS=1
# nccl env
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志
source /opt/dtk/env.sh
# hipblaslt库
export LD_LIBRARY_PATH=/data/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
# rocblas
export LD_LIBRARY_PATH=/data/blas/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1
#增加编译缓存
export cache_size_limit=64
# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b #
SAVE_PATH=./tmp_7b
TENSORBOARD_LOGS_PATH=./tmp_7b #$2 #<Specify path>
DATA_PATH="/data/datasets/oscar-1GB/oscar-1GB-llama2_text_document" #<Specify path and file prefix>_text_document
GPT_MODEL_ARGS=(
--num-layers 32
--hidden-size 4096
--ffn-hidden-size 11008
--num-attention-heads 32
--max-position-embeddings 4096
--normalization RMSNorm # LightopRMSNorm
--position-embedding-type rope # none #
--untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
)
export NVTE_FLASH_ATTN=1 # 走cutlass
# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
# --transformer-impl transformer_engine # 走core用这两组参数
# --use-mcore-models
# --transformer-impl local # 走legacy用这两组参数
# --use-legacy-models
TRAINING_ARGS=(
--transformer-impl local # 走legacy用这两组参数
--use-legacy-models
--micro-batch-size 1
--global-batch-size 256 #256 #240 #60 #512 #64
--train-iters 50
--weight-decay 0.1
--adam-beta1 0.9
--adam-beta2 0.95
--init-method-std 0.006
--clip-grad 1.0
--bf16
# --fp16 # 开启fp16需要指定loss-scale
# --loss-scale 1024
--use-distributed-optimizer
--disable-bias-linear
--attention-dropout 0
--hidden-dropout 0
# --no-gradient-accumulation-fusion
--swiglu
--lr 3.0e-5
--lr-decay-style cosine
--min-lr 3.0e-6
--lr-warmup-iters 1
--ckpt-format torch
--ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
# --recompute-granularity full # 开启重计算降低显存增加耗时
# --recompute-num-layers 5 #0 #
# --recompute-method block
--overlap-grad-reduce # 重叠ddp grad reduce
# --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
# --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠
--use-flash-attn
)
# 使用torch fa的环境变量
# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
# export TORCHINDUCTOR_BENCHMARK_FUSION=1
# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
# export TORCHINDUCTOR_MAX_AUTOTUNE=1
# export TORCHINDUCTOR_CACHE_DIR=./cache
# --use-flash-attn-cutlass # cutlass fa
# --use-flash-attn-triton # triton fa
# --use-flash-attn-torch # torch fa
MODEL_PARALLEL_ARGS=(
--sequence-parallel
--tensor-model-parallel-size 1
--pipeline-model-parallel-size 2
# --context-parallel-size 2
# --num-layers-per-virtual-pipeline-stage 4
# --microbatch-group-size-per-virtual-pipeline-stage 1
# --no-overlap-p2p-communication # 开启后
)
DATA_ARGS=(
--data-path $DATA_PATH
--seq-length 4096 #4096
--split 949,50,1
--tokenizer-type Llama2Tokenizer
--tokenizer-model /data/model_weights/llama2_7b_hf/tokenizer.model
)
EVAL_AND_LOGGING_ARGS=(
--log-throughput
--eval-iters 50
--log-interval 1
--save-interval 1000
--eval-interval 1000
--save $SAVE_PATH
--load $SAVE_PATH
--tensorboard-dir $TENSORBOARD_LOGS_PATH
)
# FINETUNE_ARGS=(
# # --finetune
# # --pretrained-checkpoint $CHECKPOINT_PATH
# --load $CHECKPOINT_PATH
# --no-load-optim
# --no-load-rng
# )
PROFILE_ARGS=(
--profile
--profile-step-start 4
--profile-step-end 5
--use-pytorch-profiler
--profile-ranks 0 1 2 3 4 5 6 7
--profile-dir prof_data
)
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=34577
DISTRIBUTED_ARGS=(
--rank ${RANK}
--world-size ${WORLD_SIZE}
--local-rank ${LOCAL_RANK}
--dist-url tcp://${DIST_URL}:${DIST_PORT}
)
APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
${GPT_MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${MODEL_PARALLEL_ARGS[@]} \
${DATA_ARGS[@]} \
${EVAL_AND_LOGGING_ARGS[@]} \
${DISTRIBUTED_ARGS[@]} \
"
# 开启profile
# ${PROFILE_ARGS[@]} \
# export HIP_VISIBLE_DEVICES=0,7 # # 4,5,6,7 #,
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # # 4,5,6,7 #,
# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
# ${APP}
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
esac
\ No newline at end of file
examples/llama/Llama3_405b.sh 0 → 100644
View file @ ce02cd51
#!/bin/bash
# set -eux
#export FLASH_ATTENTION_PRINT_PARAM=1
# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB # SYS
#export HIP_ALLOC_INITIALIZE=0
# export GPU_MAX_HW_QUEUES=10
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志
source /opt/dtk/env.sh
# 导入hipblaslt库
# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
# 更新rocblas
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1
# prof采集添加同步, 避免卡顿
# export GPU_FLUSH_ON_EXECUTION=1
# export HIP_DIRECT_DISPATCH=0
# 采集rocblas size
# export ROCBLAS_LAYER=3
# 采集 fa size
# export FLASH_ATTENTION_PRINT_PARAM=1
#增加编译缓存
export cache_size_limit=64
# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b #
SAVE_PATH=./tmp_7b
TENSORBOARD_LOGS_PATH=./tmp_7b #$2 #<Specify path>
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
GPT_MODEL_ARGS=(
--num-layers 126 #96 #8 # 126
--hidden-size 16384
--ffn-hidden-size 53248
--num-attention-heads 128
--max-position-embeddings 16384
--group-query-attention
--num-query-groups 16
--normalization RMSNorm
--position-embedding-type rope
--untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
)
export NVTE_FLASH_ATTN=1 # 走cutlass
# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
# --transformer-impl transformer_engine # 走core用这两组参数
# --use-mcore-models
# --transformer-impl local # 走legacy用这两组参数
# --use-legacy-models
TRAINING_ARGS=(
--transformer-impl transformer_engine # 走core用这两组参数
--use-mcore-models
--micro-batch-size 1
--global-batch-size 6912 # 252 #32 # 64 #240 #60 #512 #64
--train-iters 100
--weight-decay 0.1
--adam-beta1 0.9
--adam-beta2 0.95
--init-method-std 0.006
--clip-grad 1.0
--bf16
# --fp16 # 开启fp16需要指定loss-scale
# --loss-scale 1024
--use-distributed-optimizer
--disable-bias-linear
--attention-dropout 0
--hidden-dropout 0
# --no-gradient-accumulation-fusion
--swiglu
--lr 3.0e-5
--lr-decay-style cosine
--min-lr 3.0e-6
--lr-warmup-iters 1
--ckpt-format torch
--ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
# --recompute-granularity full # 开启重计算降低显存增加耗时
# --recompute-num-layers 5 #0 #
# --recompute-method block
--overlap-grad-reduce # 重叠ddp grad reduce
# --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
# --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 优化项未适配
--use-flash-attn-cutlass
)
# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
# export TORCHINDUCTOR_BENCHMARK_FUSION=1
# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
# export TORCHINDUCTOR_MAX_AUTOTUNE=1
# export TORCHINDUCTOR_CACHE_DIR=./cache
# --use-flash-attn-cutlass # cutlass fa
# --use-flash-attn-triton # triton fa
# --use-flash-attn-torch # torch fa
MODEL_PARALLEL_ARGS=(
--sequence-parallel
--tensor-model-parallel-size 8
--pipeline-model-parallel-size 18 # 7 layer/gpu
--context-parallel-size 2
)
DATA_ARGS=(
--data-path $DATA_PATH
--seq-length 4096 #4096
--split 949,50,1
--tokenizer-type Llama3Tokenizer
--tokenizer-model /public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model
# --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
)
EVAL_AND_LOGGING_ARGS=(
--log-interval 1
--log-throughput
--save-interval 1000
--eval-interval 1000
--save $SAVE_PATH
--load $SAVE_PATH
--eval-iters 10
--tensorboard-dir $TENSORBOARD_LOGS_PATH
)
# FINETUNE_ARGS=(
# # --finetune
# # --pretrained-checkpoint $CHECKPOINT_PATH
# --load $CHECKPOINT_PATH
# --no-load-optim
# --no-load-rng
# )
PROFILE_ARGS=(
--profile
--profile-step-start 4
--profile-step-end 5
--use-pytorch-profiler
--profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
--profile-dir prof_data
)
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=34577
DISTRIBUTED_ARGS=(
--rank ${RANK}
--world-size ${WORLD_SIZE}
--local-rank ${LOCAL_RANK}
--dist-url tcp://${DIST_URL}:${DIST_PORT}
)
APP="python -u pretrain_gpt.py \
${GPT_MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${MODEL_PARALLEL_ARGS[@]} \
${DATA_ARGS[@]} \
${EVAL_AND_LOGGING_ARGS[@]} \
${DISTRIBUTED_ARGS[@]} \
"
# 开启profile
# ${PROFILE_ARGS[@]} \
# export HIP_VISIBLE_DEVICES=0,7 # # 4,5,6,7 #,
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # # 4,5,6,7 #,
# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
${APP}
# case ${LOCAL_RANK} in
# [0])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# [1])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# numactl --cpunodebind=1 --membind=1 ${APP}
# ;;
# [2])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# numactl --cpunodebind=2 --membind=2 ${APP}
# ;;
# [3])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# numactl --cpunodebind=3 --membind=3 ${APP}
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# [4])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# numactl --cpunodebind=4 --membind=4 ${APP}
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# [5])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# numactl --cpunodebind=5 --membind=5 ${APP}
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# [6])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# numactl --cpunodebind=6 --membind=6 ${APP}
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# [7])
# export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# numactl --cpunodebind=7 --membind=7 ${APP}
# # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
# ;;
# esac
\ No newline at end of file
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