Support llama3 autoparallel + pipelining #1657
Conversation
meta-cla
bot
added
the
CLA Signed
wconstab
requested review from
fmassa,
ezyang,
sanketpurandare,
bdhirsh and
xmfan
and removed request for
fegin,
wwwjn and
tianyu-l
| assert parallel_dims.cp_enabled is False, "CP not supported yet" | ||
| assert parallel_dims.pp_enabled is False, "PP not supported yet" | ||
|
|
||
| pp_degree = job_config.parallelism.pipeline_parallel_degree |
There was a problem hiding this comment.
unused pp degree config, should probably raise error when its not local world size
There was a problem hiding this comment.
i deleted it (it was unused/unneeded). I don't think we need to raise any error. pp_degree does not need to equal any particular size, and pp can even be disabled.
| spmd_dims.append("tp") | ||
| spmd_mesh = world_mesh[spmd_dims] | ||
|
|
||
| dp_degree = 1 |
There was a problem hiding this comment.
same, config could specify dp_degree
There was a problem hiding this comment.
| inputs, target=targets, losses=losses, input_batch=inputs | ||
| # TODO: input_batch kwarg only needed for CP, but | ||
| # autoparallel doesn't accept kwargs in its forward | ||
| inputs, target=targets, losses=losses #, input_batch=inputs |
There was a problem hiding this comment.
Curious, why does CP need input_batch?
There was a problem hiding this comment.
I assumed you would know. Am I wrong?
There was a problem hiding this comment.
oh, there was a change to remove the need of input_batch. We may want to do the same change to autoparalle.
|
|
||
| pp_degree = job_config.parallelism.pipeline_parallel_degree | ||
| local_batch_size = job_config.training.local_batch_size | ||
| spmd_batch_size = local_batch_size |
There was a problem hiding this comment.
oops this is a bug for the non-pp case. should be local *dp degree and put in an 'else' branch
| self.pp_schedule.step( | ||
| inputs, target=targets, losses=losses, input_batch=inputs | ||
| # TODO: input_batch kwarg only needed for CP, but | ||
| # autoparallel doesn't accept kwargs in its forward |
| # # step. | ||
| # dp_degree = parallel_dims.dp_replicate * parallel_dims.dp_shard | ||
| # global_batch_size = job_config.training.local_batch_size * dp_degree | ||
| if parallel_dims.pp_enabled and pp_rank > 0: |
There was a problem hiding this comment.
What a mess. No action here needed, but it's definitely worth thinking about what the terminal UX state here should be.
so far just tested locally `LOG_RANK=4 CONFIG_FILE=././torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name llama3_auto_parallel --parallelism.pipeline_parallel_degree 2 --training.steps 100` Runs and loss converges. Left one TODO about global-batch-size and gradient accumulation
wconstab
force-pushed
the
whc/autop_pipeline
branch
from
7e60a61 to
188b002
Compare
so far just tested locally
LOG_RANK=4 CONFIG_FILE=././torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name llama3_auto_parallel --parallelism.pipeline_parallel_degree 2 --training.steps 100Runs and loss converges.
Left one TODO about global-batch-size and gradient accumulation