github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Merge pull request #6114 from duanjunwen/dev/zero_bubble 

[Zerobubble] Support LinearWithAsyncCommunication for sharderformer policy
feature/zerobubble
duanjunwen 3 days ago committed by GitHub
parent
37b23e32b1 41fdd2139b
commit
810cafb2f9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
12 changed files with 465 additions and 132 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 1
      colossalai/pipeline/p2p.py
  2. 150
      colossalai/pipeline/schedule/zero_bubble_pp.py
  3. 23
      colossalai/shardformer/layer/attn.py
  4. 1
      colossalai/shardformer/modeling/llama.py
  5. 1
      colossalai/shardformer/modeling/mixtral.py
  6. 98
      colossalai/shardformer/policies/bert.py
  7. 93
      colossalai/shardformer/policies/llama.py
  8. 71
      colossalai/shardformer/policies/mistral.py
  9. 76
      colossalai/shardformer/policies/mixtral.py
  10. 7
      examples/language/llama/benchmark.py
  11. 2
      examples/language/mixtral/benchmark.py
  12. 32
      tests/test_pipeline/test_schedule/test_zerobubble_pp.py

1
colossalai/pipeline/p2p.py
Unescape View File

@ -432,7 +432,6 @@ def _communicate(
overlap_p2p=overlap_p2p,
send_first=send_first if send_first != None else True,
)
if metadata_recv is not None:
assert isinstance(metadata_recv, P2PMetadata)
tree_spec = metadata_recv.tree_spec

150
colossalai/pipeline/schedule/zero_bubble_pp.py
Unescape View File

@ -64,10 +64,28 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# P2PMeta cache
self.enable_metadata_cache = enable_metadata_cache
self.send_tensor_metadata = True
self.send_grad_metadata = True
self.tensor_metadata_recv = None
self.grad_metadata_recv = None
# check send_tensor_metadata, send_grad_metadata
# pp4 as sample, we should follow this meta strategy
# send_tensor_meta(fwd) send_grad_meta(bwd)
# chunk0 | chunk1 chunk0 | chunk 1
# stage 0 T | F F | T
# stage 1 T | T T | T
# stage 2 T | T T | T
# stage 3 F | T F | T
if stage_manager.is_first_stage(ignore_chunk=True):
self.send_tensor_metadata = [True, False]
self.send_grad_metadata = [False, True]
elif stage_manager.is_last_stage(ignore_chunk=True):
self.send_tensor_metadata = [False, True]
self.send_grad_metadata = [True, False]
else:
self.send_tensor_metadata = [True, True]
self.send_grad_metadata = [True, True]
# meta cache buffer
self.tensor_metadata_recv = [None, None] # [chunk 0 meta, chunk 1 meta]
self.grad_metadata_recv = [None, None]
# P2P communication
self.comm = PipelineP2PCommunication(stage_manager, overlap_p2p=overlap_p2p)
@ -96,10 +114,16 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
self.output_tensors_grad_dw = [[], []]
# buffer for communication
self.send_forward_buffer = [[], []]
self.recv_forward_buffer = [[], []]
self.send_backward_buffer = [[], []]
self.recv_backward_buffer = [[], []]
self.send_forward_buffer = [[], []] # [chunk0:[torch.Tensor], chunk1:[torch.Tensor]]
self.recv_forward_buffer = [
[],
[],
] # [chunk0:[(torch.Tensor, wait_handle)], chunk1:[(torch.Tensor, wait_handle)]]
self.send_backward_buffer = [[], []] # [chunk0:[torch.Tensor], chunk1:[torch.Tensor]]
self.recv_backward_buffer = [
[],
[],
] # [chunk0:[(torch.Tensor, wait_handle)], chunk1:[(torch.Tensor, wait_handle)]]
# y buffer for local send fwd
self.local_send_forward_buffer = []
@ -225,7 +249,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; cause u are chunk 0 in first rank, u have no prev rank;
#################
if self.stage_manager.is_first_stage(ignore_chunk=True):
# return None, []
return []
################
@ -235,12 +258,11 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
else:
prev_rank = self.stage_manager.get_prev_rank()
input_tensor, wait_handles = self.comm.recv_forward(
prev_rank=prev_rank, metadata_recv=self.tensor_metadata_recv
prev_rank=prev_rank, metadata_recv=self.tensor_metadata_recv[model_chunk_id]
)
if self.enable_metadata_cache and self.tensor_metadata_recv is None:
self.tensor_metadata_recv = create_send_metadata(input_tensor)
self.recv_forward_buffer[model_chunk_id].append(input_tensor)
# return input_tensor, wait_handles
if self.enable_metadata_cache and self.tensor_metadata_recv[model_chunk_id] is None:
self.tensor_metadata_recv[model_chunk_id] = create_send_metadata(input_tensor)
self.recv_forward_buffer[model_chunk_id].append((input_tensor, wait_handles))
return wait_handles
else:
@ -259,12 +281,11 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
else:
next_rank = self.stage_manager.get_next_rank()
input_tensor, wait_handles = self.comm.recv_forward(
next_rank, metadata_recv=self.tensor_metadata_recv
next_rank, metadata_recv=self.tensor_metadata_recv[model_chunk_id]
)
if self.enable_metadata_cache and self.tensor_metadata_recv is None:
self.tensor_metadata_recv = create_send_metadata(input_tensor)
self.recv_forward_buffer[model_chunk_id].append(input_tensor)
# return input_tensor, wait_handles
if self.enable_metadata_cache and self.tensor_metadata_recv[model_chunk_id] is None:
self.tensor_metadata_recv[model_chunk_id] = create_send_metadata(input_tensor)
self.recv_forward_buffer[model_chunk_id].append((input_tensor, wait_handles))
return wait_handles
def recv_backward(self, model_chunk_id: int, next_rank: int = None) -> List:
@ -287,7 +308,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; Already get dy from local_send_backward_buffer in schedule b
################
if self.stage_manager.is_last_stage(ignore_chunk=True):
# return None, []
return []
################
@ -297,12 +317,11 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
else:
next_rank = self.stage_manager.get_next_rank()
output_tensor_grad, wait_handles = self.comm.recv_backward(
next_rank, metadata_recv=self.grad_metadata_recv
next_rank, metadata_recv=self.grad_metadata_recv[model_chunk_id]
)
if self.enable_metadata_cache and self.grad_metadata_recv is None:
self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
self.recv_backward_buffer[model_chunk_id].append(output_tensor_grad)
# return output_tensor_grad, wait_handles
if self.enable_metadata_cache and self.grad_metadata_recv[model_chunk_id] is None:
self.grad_metadata_recv[model_chunk_id] = create_send_metadata(output_tensor_grad)
self.recv_backward_buffer[model_chunk_id].append((output_tensor_grad, wait_handles))
return wait_handles
else:
@ -312,7 +331,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; get loss from local
################
if self.stage_manager.is_first_stage(ignore_chunk=True):
# return None, []
return []
################
@ -322,12 +340,11 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
else:
prev_rank = self.stage_manager.get_prev_rank()
output_tensor_grad, wait_handles = self.comm.recv_backward(
next_rank=prev_rank, metadata_recv=self.grad_metadata_recv
next_rank=prev_rank, metadata_recv=self.grad_metadata_recv[model_chunk_id]
)
if self.enable_metadata_cache and self.grad_metadata_recv is None:
self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
self.recv_backward_buffer[model_chunk_id].append(output_tensor_grad)
# return output_tensor_grad, wait_handles
if self.enable_metadata_cache and self.grad_metadata_recv[model_chunk_id] is None:
self.grad_metadata_recv[model_chunk_id] = create_send_metadata(output_tensor_grad)
self.recv_backward_buffer[model_chunk_id].append((output_tensor_grad, wait_handles))
return wait_handles
def send_forward(self, model_chunk_id: int, next_rank: int = None) -> List:
@ -349,6 +366,7 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; hold y on local_send_forward_buffer
################
if self.stage_manager.is_last_stage(ignore_chunk=True):
self.send_tensor_metadata[model_chunk_id] = not self.enable_metadata_cache
return []
################
@ -359,9 +377,11 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
next_rank = self.stage_manager.get_next_rank()
output_tensor = self.send_forward_buffer[model_chunk_id].pop(0)
send_handles = self.comm.send_forward(
output_object=output_tensor, next_rank=next_rank, send_metadata=self.send_tensor_metadata
output_object=output_tensor,
next_rank=next_rank,
send_metadata=self.send_tensor_metadata[model_chunk_id],
)
self.send_tensor_metadata = not self.enable_metadata_cache
self.send_tensor_metadata[model_chunk_id] = not self.enable_metadata_cache
return send_handles
else:
@ -370,6 +390,7 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; Already send LOSS to local_send_backward_buffer in schedule f send part
################
if self.stage_manager.is_first_stage(ignore_chunk=True):
self.send_tensor_metadata[model_chunk_id] = not self.enable_metadata_cache
return []
################
@ -380,9 +401,9 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
prev_rank = self.stage_manager.get_prev_rank()
output_tensor = self.send_forward_buffer[model_chunk_id].pop(0)
send_handles = self.comm.send_forward(
output_tensor, prev_rank, send_metadata=self.send_tensor_metadata
output_tensor, prev_rank, send_metadata=self.send_tensor_metadata[model_chunk_id]
)
self.send_tensor_metadata = not self.enable_metadata_cache
self.send_tensor_metadata[model_chunk_id] = not self.enable_metadata_cache
return send_handles
def send_backward(self, model_chunk_id: int, prev_rank: int = None) -> List:
@ -405,6 +426,7 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; cause u are the first chunk in first stage; bwd end
################
if self.stage_manager.is_first_stage(ignore_chunk=True):
self.send_grad_metadata[model_chunk_id] = not self.enable_metadata_cache
return []
################
@ -415,9 +437,9 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
prev_rank = self.stage_manager.get_prev_rank()
input_tensor_grad = self.send_backward_buffer[model_chunk_id].pop(0)
send_handles = self.comm.send_backward(
input_tensor_grad, prev_rank, send_metadata=self.send_grad_metadata
input_tensor_grad, prev_rank, send_metadata=self.send_grad_metadata[model_chunk_id]
)
self.send_grad_metadata = not self.enable_metadata_cache
self.send_grad_metadata[model_chunk_id] = not self.enable_metadata_cache
return send_handles
# bwd chunk1 is left V;
@ -427,6 +449,7 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# do nothing; Already send input_tensor_grad to local_send_bwd_buffer in schedule b;
################
if self.stage_manager.is_last_stage(ignore_chunk=True):
self.send_grad_metadata[model_chunk_id] = not self.enable_metadata_cache
return []
################
@ -437,9 +460,9 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
next_rank = self.stage_manager.get_next_rank()
input_tensor_grad = self.send_backward_buffer[model_chunk_id].pop(0)
send_handles = self.comm.send_backward(
input_tensor_grad, next_rank, send_metadata=self.send_grad_metadata
input_tensor_grad, next_rank, send_metadata=self.send_grad_metadata[model_chunk_id]
)
self.send_grad_metadata = not self.enable_metadata_cache
self.send_grad_metadata[model_chunk_id] = not self.enable_metadata_cache
return send_handles
def forward_step(
@ -519,8 +542,6 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
output_obj_grad_ = []
# For chunk 0 stage 0, use micro_batch as input_obj_; and we don't have to cal microbatch dx.
# if model_chunk_id == 0 and self.stage_manager.is_first_stage(ignore_chunk=True):
# return None
# For loss backward; output_obj is loss; output_obj_grad should be None
if model_chunk_id == 1 and self.stage_manager.is_first_stage(ignore_chunk=True):
@ -633,9 +654,12 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
if model_chunk_id == 0:
# is first stage; get input from microbatch
if self.stage_manager.is_first_stage(ignore_chunk=True):
input_obj = None
input_obj = None # (tensor, wait_handle)
else:
input_obj = self.recv_forward_buffer[model_chunk_id].pop(0)
for h in input_obj[1]:
h.wait()
input_obj = input_obj[0]
else:
# is last stage; recv from local
if self.stage_manager.is_last_stage(ignore_chunk=True):
@ -643,7 +667,9 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# not last stage; recv from next
else:
input_obj = self.recv_forward_buffer[model_chunk_id].pop(0)
for h in input_obj[1]:
h.wait()
input_obj = input_obj[0]
# Here, let input_obj.requires_grad_()
# if input_obj is not None:
if not isinstance(input_obj, torch.Tensor):
@ -689,10 +715,8 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# Do not release_tensor_data loss, release_tensor_data other output_obj;
if model_chunk_id == 1 and self.stage_manager.is_first_stage(ignore_chunk=True):
self.output_tensors[model_chunk_id].append(output_obj)
# self.output_tensors_dw[model_chunk_id].append(output_obj)
else:
self.output_tensors[model_chunk_id].append(output_obj)
# self.output_tensors_dw[model_chunk_id].append(output_obj)
# add output to send_fwd_buffer
if model_chunk_id == 0: # chunk 0
@ -732,6 +756,9 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# chunk0 not last stage; recv output_grad from recv_backward_buffer
else:
output_tensor_grad = self.recv_backward_buffer[model_chunk_id].pop(0)
for h in output_tensor_grad[1]:
h.wait()
output_tensor_grad = output_tensor_grad[0]
else:
# chunk1, is first stage; recv LOSS from local send bwd buffer
if self.stage_manager.is_first_stage(ignore_chunk=True):
@ -739,25 +766,14 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# chunk1, not first stage; recv output_grad from recv_backward_buffer
else:
output_tensor_grad = self.recv_backward_buffer[model_chunk_id].pop(0)
for h in output_tensor_grad[1]:
h.wait()
output_tensor_grad = output_tensor_grad[0]
# get input and output object from buffer;
input_obj = self.input_tensors[model_chunk_id].pop(0)
output_obj = self.output_tensors[model_chunk_id].pop(0)
# # save output_tensor_grad for dw
# if model_chunk_id == 1 and self.stage_manager.is_first_stage(ignore_chunk=True):
# # we save loss here
# self.output_tensors_grad_dw[model_chunk_id].append(output_obj)
# else:
# # we save output_tensor_grad here
# self.output_tensors_grad_dw[model_chunk_id].append(output_tensor_grad)
# the_output_obj_grad = []
# if isinstance(output_obj, dict):
# for (k, v) in output_obj.items():
# the_output_obj_grad.append(v.requires_grad)
# else:
# the_output_obj_grad.append(output_obj.requires_grad)
input_object_grad = self.backward_b_step(
model_chunk=model_chunk,
model_chunk_id=model_chunk_id,
@ -800,20 +816,8 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
Returns:
Nothing.
"""
# get y & dy from buffer
# output_obj = self.output_tensors_dw[model_chunk_id].pop(0)
# output_obj_grad = self.output_tensors_grad_dw[model_chunk_id].pop(0)
WeightGradStore.pop(chunk=model_chunk_id)
# self.backward_w_step(
# model_chunk=model_chunk,
# model_chunk_id=model_chunk_id,
# optimizer=optimizer,
# output_obj=output_obj,
# output_obj_grad=output_obj_grad,
# )
def run_forward_only(
self,
model_chunk: Union[ModuleList, Module],
@ -890,6 +894,8 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
# communication
communication_func = self.communication_map[scheduled_node.type]
wait_handle = communication_func(scheduled_node.chunk)
# We wait recv handle in fwd step and bwd step. Here only need to wait for send handle
if scheduled_node.type in {"SEND_FORWARD", "SEND_BACKWARD"}:
self.wait_handles.append(wait_handle)
elif scheduled_node.type == "F":
self.schedule_f(
@ -914,10 +920,10 @@ class ZeroBubbleVPipeScheduler(PipelineSchedule):
model_chunk_id=scheduled_node.chunk,
optimizer=optimizer,
)
# wait here to ensure all communication is done
for h in self.wait_handles:
for hh in h:
hh.wait()
# return loss & output
if outputs is not None:
outputs = merge_batch(outputs)

23
colossalai/shardformer/layer/attn.py
Unescape View File

@ -6,6 +6,7 @@ import torch.distributed
import torch.distributed as dist
import torch.nn.functional as F
from einops import rearrange
from packaging import version
from colossalai.kernel.kernel_loader import (
FlashAttentionDaoLoader,
@ -642,9 +643,7 @@ class RingAttention(torch.autograd.Function):
max_seqlen_q = max_seqlen_kv = max_seqlen
cu_seqlens_half = cu_seqlens // 2
max_seqlen_half = max_seqlen // 2
misc_kwargs = {
"window_size": (-1, -1),
"alibi_slopes": None,
"softmax_scale": q.shape[-1] ** -0.5 if softmax_scale is None else softmax_scale,
"dropout_p": dropout_p,
@ -652,6 +651,13 @@ class RingAttention(torch.autograd.Function):
"softcap": 0.0,
"return_softmax": False,
}
import flash_attn
if version.parse(flash_attn.__version__) > version.parse("2.6.3"):
misc_kwargs["window_size_left"] = -1
misc_kwargs["window_size_right"] = -1
else:
misc_kwargs["window_size"] = (-1, -1)
if (
RingAttention.HALF_INDICES is not None
@ -707,6 +713,19 @@ class RingAttention(torch.autograd.Function):
# Helper to pass args to FA
def _forward(q, k, v, causal):
if version.parse(flash_attn.__version__) > version.parse("2.6.3"):
(out, softmax_lse, S_dmask, rng_state) = _flash_attn_forward(
q,
k,
v,
cu_seqlens_q if q.shape[0] == t else cu_seqlens_half,
cu_seqlens_kv if k.shape[0] == t else cu_seqlens_half,
max_seqlen_q if q.shape[0] == t else max_seqlen_half,
max_seqlen_kv if k.shape[0] == t else max_seqlen_half,
causal=causal,
**misc_kwargs,
)
else:
(
_,
_,

1
colossalai/shardformer/modeling/llama.py
Unescape View File

@ -191,7 +191,6 @@ class LlamaPipelineForwards:
num_model_chunks=stage_manager.num_model_chunks,
)
assert num_ckpt_layers <= end_idx - start_idx
for idx, decoder_layer in enumerate(self.layers[start_idx:end_idx], start=start_idx):
if output_hidden_states:
all_hidden_states += (hidden_states,)

1
colossalai/shardformer/modeling/mixtral.py
Unescape View File

@ -381,7 +381,6 @@ class MixtralPipelineForwards:
output_router_logits,
use_cache,
)
hidden_states = layer_outputs[0]
if use_cache:

98
colossalai/shardformer/policies/bert.py
Unescape View File

@ -75,6 +75,8 @@ class BertPolicy(Policy):
sp_partial_derived = sp_mode == "split_gather"
use_zbv = self.pipeline_stage_manager is not None and self.pipeline_stage_manager.use_zbv
if self.shard_config.enable_tensor_parallelism:
assert (
self.model.config.num_attention_heads % self.shard_config.tensor_parallel_size == 0
@ -97,6 +99,7 @@ class BertPolicy(Policy):
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -105,6 +108,7 @@ class BertPolicy(Policy):
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -113,6 +117,7 @@ class BertPolicy(Policy):
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -125,6 +130,7 @@ class BertPolicy(Policy):
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -138,6 +144,7 @@ class BertPolicy(Policy):
"seq_parallel_mode": sp_mode,
"skip_bias_add": self.enable_bias_gelu_fused,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -146,6 +153,97 @@ class BertPolicy(Policy):
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="output.dropout",
target_module=col_nn.DropoutForParallelInput,
),
],
)
policy[BertEmbeddings] = ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="dropout",
target_module=col_nn.DropoutForReplicatedInput,
),
]
)
if self.enable_bias_gelu_fused:
self.append_or_create_method_replacement(
description={
"forward": get_jit_fused_bert_intermediate_forward(),
},
policy=policy,
target_key=BertIntermediate,
)
elif use_zbv:
policy[BertLayer] = ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="attention.self.query",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.self.key",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.self.value",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.self.dropout",
target_module=col_nn.DropoutForParallelInput,
),
SubModuleReplacementDescription(
suffix="attention.output.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="attention.output.dropout",
target_module=col_nn.DropoutForParallelInput,
),
SubModuleReplacementDescription(
suffix="intermediate.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"skip_bias_add": self.enable_bias_gelu_fused,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="output.dense",
target_module=col_nn.LinearWithGradAccum,
kwargs={
"seq_parallel_mode": sp_mode,
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(

93
colossalai/shardformer/policies/llama.py
Unescape View File

@ -9,6 +9,7 @@ from colossalai.shardformer.layer import (
FusedRMSNorm,
Linear1D_Col,
Linear1D_Row,
LinearWithGradAccum,
PaddingEmbedding,
PaddingLMHead,
RMSNorm,
@ -104,7 +105,7 @@ class LlamaPolicy(Policy):
policy=policy,
target_key=LlamaModel,
)
# enable tp, replace layer to tp Linear1D_Col,Linear1D_Row,
if self.shard_config.enable_tensor_parallelism:
assert (
num_q_heads % tp_size == 0
@ -191,6 +192,76 @@ class LlamaPolicy(Policy):
],
)
# not enable tp, replace layer to LinearWithGradAccum
elif use_zbv:
policy[LlamaDecoderLayer] = ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="self_attn.q_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="self_attn.k_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="self_attn.v_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="self_attn.o_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="mlp.gate_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="mlp.up_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
SubModuleReplacementDescription(
suffix="mlp.down_proj",
target_module=LinearWithGradAccum,
kwargs=dict(
seq_parallel_mode=sp_mode,
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
),
],
)
if embedding_cls is not None:
self.append_or_create_submodule_replacement(
description=SubModuleReplacementDescription(
@ -416,6 +487,7 @@ class LlamaForSequenceClassificationPolicy(LlamaPolicy):
policy = super().module_policy()
use_zbv = self.pipeline_stage_manager is not None and self.pipeline_stage_manager.use_zbv
# enable tp, replace layer to tp Linear1D_Col,Linear1D_Row,
if self.shard_config.enable_tensor_parallelism:
# add a new item for sequence classification
new_item = {
@ -434,6 +506,25 @@ class LlamaForSequenceClassificationPolicy(LlamaPolicy):
)
}
policy.update(new_item)
# enable tp, replace layer to LinearWithGradAccum
elif use_zbv:
# add a new item for sequence classification
new_item = {
LlamaForSequenceClassification: ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="score",
target_module=LinearWithGradAccum,
kwargs=dict(
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
)
]
)
}
policy.update(new_item)
# to be confirmed
if self.pipeline_stage_manager:
# set None as default

71
colossalai/shardformer/policies/mistral.py
Unescape View File

@ -10,6 +10,7 @@ from colossalai.shardformer.layer import (
FusedRMSNorm,
Linear1D_Col,
Linear1D_Row,
LinearWithGradAccum,
PaddingEmbedding,
PaddingLMHead,
VocabParallelEmbedding1D,
@ -62,6 +63,8 @@ class MistralPolicy(Policy):
if self.tie_weight:
embedding_cls = PaddingEmbedding
use_zbv = self.pipeline_stage_manager is not None and self.pipeline_stage_manager.use_zbv
if self.shard_config.enable_sequence_parallelism:
self.shard_config.enable_sequence_parallelism = False
warnings.warn(
@ -90,6 +93,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -97,6 +101,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -104,6 +109,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -111,6 +117,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -118,6 +125,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -125,6 +133,7 @@ class MistralPolicy(Policy):
target_module=Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
@ -132,6 +141,68 @@ class MistralPolicy(Policy):
target_module=Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
],
)
elif use_zbv:
policy[MistralDecoderLayer] = ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="self_attn.q_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.k_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.v_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.o_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="mlp.gate_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="mlp.up_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="mlp.down_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
],

76
colossalai/shardformer/policies/mixtral.py
Unescape View File

@ -7,9 +7,18 @@ from torch import Tensor
from torch.nn import Module
from transformers.models.mixtral.modeling_mixtral import MixtralForCausalLM, MixtralModel
from colossalai.shardformer.layer import FusedRMSNorm, Linear1D_Col
from colossalai.shardformer.layer.embedding import PaddingEmbedding, VocabParallelEmbedding1D
from colossalai.shardformer.layer.linear import Linear1D_Row
from colossalai.shardformer.layer import (
FusedRMSNorm,
Linear1D_Col,
Linear1D_Row,
LinearWithGradAccum,
PaddingEmbedding,
VocabParallelEmbedding1D,
)
# from colossalai.shardformer.layer import FusedRMSNorm, Linear1D_Col
# from colossalai.shardformer.layer.embedding import PaddingEmbedding, VocabParallelEmbedding1D
# from colossalai.shardformer.layer.linear import Linear1D_Row
from colossalai.shardformer.modeling.mixtral import (
EPMixtralSparseMoeBlock,
MixtralPipelineForwards,
@ -166,6 +175,51 @@ class MixtralPolicy(Policy):
],
)
elif use_zbv:
policy[MixtralDecoderLayer] = ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="self_attn.q_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.k_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.v_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="self_attn.o_proj",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
SubModuleReplacementDescription(
suffix="block_sparse_moe.gate",
target_module=LinearWithGradAccum,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
"use_zbv": use_zbv,
},
),
],
)
if embedding_cls is not None:
self.append_or_create_submodule_replacement(
description=SubModuleReplacementDescription(
@ -351,6 +405,22 @@ class MixtralForCausalLMPolicy(MixtralPolicy):
)
}
policy.update(new_item)
elif use_zbv:
new_item = {
MixtralForCausalLM: ModulePolicyDescription(
sub_module_replacement=[
SubModuleReplacementDescription(
suffix="lm_head",
target_module=LinearWithGradAccum,
kwargs=dict(
fp8_communication=self.shard_config.fp8_communication,
use_zbv=use_zbv,
),
)
],
)
}
policy.update(new_item)
if self.pipeline_stage_manager:
# set None as default

7
examples/language/llama/benchmark.py
Unescape View File

@ -163,8 +163,6 @@ def main():
enable_async_reduce=not args.disable_async_reduce,
use_fp8=args.use_fp8,
fp8_communication=args.use_fp8_comm,
use_fp8=args.use_fp8,
fp8_communication=args.use_fp8_comm,
)
elif args.plugin == "gemini_auto":
plugin = GeminiPlugin(
@ -179,8 +177,6 @@ def main():
enable_flash_attention=args.xformers,
use_fp8=args.use_fp8,
fp8_communication=args.use_fp8_comm,
use_fp8=args.use_fp8,
fp8_communication=args.use_fp8_comm,
)
elif args.plugin == "fsdp":
if use_empty_init:
@ -192,7 +188,6 @@ def main():
),
param_init_fn=empty_init(),
fp8_communication=args.use_fp8_comm,
fp8_communication=args.use_fp8_comm,
)
else:
plugin = TorchFSDPPlugin(
@ -214,7 +209,6 @@ def main():
cpu_offload=CPUOffload(offload_params=True),
param_init_fn=empty_init(),
fp8_communication=args.use_fp8_comm,
fp8_communication=args.use_fp8_comm,
)
else:
plugin = TorchFSDPPlugin(
@ -225,7 +219,6 @@ def main():
),
cpu_offload=CPUOffload(offload_params=True),
fp8_communication=args.use_fp8_comm,
fp8_communication=args.use_fp8_comm,
)
elif args.plugin == "3d":
if args.pp_style == "zbv":

2
examples/language/mixtral/benchmark.py
Unescape View File

@ -122,7 +122,7 @@ def main():
num_ckpt_layers_per_stage=[19, 19, 19, 13],
),
"num_layers_per_stage": [19, 20, 20, 21],
# "pp_style": "interleaved",
"pp_style": "interleaved",
}
if args.custom_ckpt
else {}

32
tests/test_pipeline/test_schedule/test_zerobubble_pp.py
Unescape View File

@ -749,24 +749,17 @@ def run_fwd_bwd_vschedule_with_optim(test_config):
assert_optim_param_groups(optim_base_param_groups, optim_pp_param_groups)
# TODO:3) support booster & Hybrid base 2)
def run_with_hybridplugin(test_config):
pass
# TODO:4) support booster & MoEHybrid base 2)
@parameterize(
"config",
[
# (0, 1, 4, 1, 1),
# (1, 2, 2, 1, 1),
(1, 2, 1, 1, 2),
(1, 1, 2, 2, 1),
# (1, 2, 1, 2, 1),
# (1, 2, 1, 1, 2),
(1, 2, 1, 2, 1),
(1, 2, 2, 1, 1),
(1, 1, 4, 1, 1),
],
)
def run_with_booster_moehybridplugin(config: Tuple[int, ...]):
test_config = config
stage, ep_size, pp_size, tp_size, sp_size = config
num_microbatches = pp_size
dist.get_world_size()
@ -876,7 +869,6 @@ def run_with_booster_moehybridplugin(config: Tuple[int, ...]):
return_outputs=True,
)
# stage 0 chunk 0
parallel_output = None
if (
booster.plugin.stage_manager.is_first_stage(ignore_chunk=True)
and rank == dist.get_process_group_ranks(plugin.pp_group)[0]
@ -910,9 +902,7 @@ def run_with_booster_moehybridplugin(config: Tuple[int, ...]):
p.grad /= dp_size
torch_optimizer.step()
torch_optimizer.zero_grad()
assert_loose_close(parallel_output, torch_output_sum, dtype=dtype)
print(f"rank {dist.get_rank()} config {test_config} test passed")
clear_layout_converter()
Randomizer.reset_index()
torch.cuda.empty_cache()
@ -921,11 +911,11 @@ def run_with_booster_moehybridplugin(config: Tuple[int, ...]):
@parameterize(
"config",
[
(1, 2, 2, 1), # Pass
# TODO: only support pp + tp accleration; Will support fully pp and None tp Hybrid in furture;
# (0, 4, 1, 1),
# (1, 2, 1, 2),
# (1, 1, 2, 2),
# Pass
(1, 2, 2, 1),
(1, 2, 1, 2),
(1, 1, 2, 2),
(1, 4, 1, 1),
],
)
def run_with_booster_hybridplugin(config: Tuple[int, ...]):
@ -1034,7 +1024,6 @@ def run_with_booster_hybridplugin(config: Tuple[int, ...]):
return_outputs=True,
)
# stage 0 chunk 0
parallel_output = None
if (
booster.plugin.stage_manager.is_first_stage(ignore_chunk=True)
and rank == dist.get_process_group_ranks(plugin.pp_group)[0]
@ -1068,9 +1057,8 @@ def run_with_booster_hybridplugin(config: Tuple[int, ...]):
p.grad /= dp_size
torch_optimizer.step()
torch_optimizer.zero_grad()
assert_loose_close(parallel_output, torch_output_sum, dtype=dtype)
print(f"rank {dist.get_rank()} pp_size:{pp_size}, tp_size {tp_size}, sp_size :{sp_size} test passed")
clear_layout_converter()
Randomizer.reset_index()
torch.cuda.empty_cache()

Write Preview
Loading…
Cancel
Save