Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add MLP/lm_head tp grain size setting. #6828

Open
wants to merge 6 commits into
base: master
Choose a base branch
from
+15 −4
Open

Add MLP/lm_head tp grain size setting. #6828

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
55c3ddc
add tp grain size
Yejing-Lai Nov 28, 2024
6befb00
Merge branch 'master' into lyj/tp_grain_size
loadams Dec 6, 2024
a5e68e0
Merge branch 'master' into lyj/tp_grain_size
loadams Dec 10, 2024
c884621
Merge branch 'master' into lyj/tp_grain_size
loadams Dec 10, 2024
1842815
updated tp_grain_size to tensor_parallel field
Yejing-Lai Dec 11, 2024
9565d3a
Merge branch 'master' into lyj/tp_grain_size
tjruwase Dec 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions deepspeed/inference/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,6 +289,8 @@ class DeepSpeedInferenceConfig(DeepSpeedConfigModel):
"deprecated": True,
"new_param": "moe.type"
})
tp_grain_size: int = 64
"Desired MLP/lm_head tp size granularity. DNN library favors tensor size in granularity of power of 2, we pick 64 as a default size."

@field_validator("dtype", mode="before")
def validate_dtype(cls, field_value, values):
Expand Down
5 changes: 4 additions & 1 deletion deepspeed/module_inject/replace_module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@
from .layers import TensorParallelOcShardConv2d, TensorParallelIcShardConv2d

from deepspeed import comm as dist
from deepspeed.module_inject.tp_shard import set_num_kv_heads, set_n_embd, set_num_attention_heads
from deepspeed.module_inject.tp_shard import set_num_kv_heads, set_n_embd, set_num_attention_heads, set_tp_grain_size

from .load_checkpoint import load_model_with_checkpoint
import time
Expand Down Expand Up @@ -303,6 +303,9 @@ def replace_wo_policy(module, all_reduce_linears, prefix="", state_dict=None):
if hasattr(model_config, 'num_attention_heads'):
set_num_attention_heads(getattr(model_config, 'num_attention_heads'))

# 4.4 set tp_grain_size
set_tp_grain_size(config.tp_grain_size)

# 5. Set linear policies
_autotp.update_linear_policies()

Expand Down
11 changes: 8 additions & 3 deletions deepspeed/module_inject/tp_shard.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,11 @@ def set_n_embd(num):
n_embd = num


def set_tp_grain_size(num):
global tp_grain_size
tp_grain_size = num


def get_num_kv_heads():
global num_kv_heads
if 'num_kv_heads' in globals():
Expand All @@ -45,9 +50,9 @@ def get_shard_size(total_size, mp_size, name=None, rank=None):
my_slices = (num_kv_heads // mp_size) + (1 if rank < (num_kv_heads % mp_size) else 0)
return total_size * my_slices // num_kv_heads
else:
if total_size >= 64:
grain_size = total_size // 64
return (grain_size // mp_size + (1 if rank < (grain_size % mp_size) else 0)) * 64
if total_size >= tp_grain_size:
grain_size = total_size // tp_grain_size
return (grain_size // mp_size + (1 if rank < (grain_size % mp_size) else 0)) * tp_grain_size
else:
return total_size // mp_size + (1 if rank < (total_size % mp_size) else 0)

Expand Down