[QEff Finetune]: Enable PP+DDP #394
Draft
+70
−16
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Mamta Singh <[email protected]>
quic-mamta
requested review from
quic-rishinr,
ochougul and
quic-amitraj
as code owners
quic-mamta
force-pushed
the
pp_ddp
branch
2 times, most recently
from
e8b1da7 to
df36ae1
Compare
quic-mamta
force-pushed
the
pp_ddp
branch
8 times, most recently
from
3ca1229 to
53ff3c4
Compare
quic-meetkuma
suggested changes
May 12, 2025
QEfficient/cloud/finetune.py
Outdated
| - This device map structure is verified for llama models only. | ||
| """ | ||
| device_map = { | ||
| "model.embed_tokens": rank * num_pp_stages, |
There was a problem hiding this comment.
Please add some explanation why these particular layers are mapped to a particular device.
L64 to L67
QEfficient/cloud/finetune.py
Outdated
| "model.rotary_emb": rank * num_pp_stages + (num_pp_stages - 1), | ||
| } | ||
| n_layer_per_stage = math.ceil(num_layers / num_pp_stages) | ||
| for j in range(num_pp_stages): |
There was a problem hiding this comment.
Please add some strong documentation for this double for loop. It is difficult to understand without taking a case. Better add some example and explain with it.
Signed-off-by: Mamta Singh <[email protected]>
Signed-off-by: Mamta Singh <[email protected]>
quic-swatia
reviewed
Jun 2, 2025
QEfficient/cloud/finetune.py
Outdated
| n_layer_per_stage = math.ceil(num_layers / num_pp_stages) | ||
| for j in range(num_pp_stages): | ||
| for i in range(n_layer_per_stage * j, min(n_layer_per_stage * (j + 1), num_layers)): | ||
| device_map[f"model.layers.{i}"] = rank * num_pp_stages + j |
There was a problem hiding this comment.
Will this not place 2 extra layers than n_layer_per_stage on the first device?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Added support for PP and DDP
Command for PP only : QAIC_VISIBLE_DEVICES=0,1,2,3 python -m QEfficient.cloud.finetune --device qaic --enable_pp --dist_backend qccl (number of pipeline stages will be equal to visible devices)
Command for DDP only : QAIC_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node 4 -m QEfficient.cloud.finetune --device qaic --enable_ddp --dist_backend qccl
Command for PP+DDP : For 4 qaic devices(1 Ultra) with 2 pipeline stages
QAIC_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc-per-node 2 -m QEfficient.cloud.finetune --device qaic --enable_ddp --enable_pp --num_pp_stages 2 --dist_backend qccl