Enable 2D sharding #17
Merged
Conversation
This file contains 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
jonb377
approved these changes
Aug 1, 2023
Comment on lines
540
to
541
| data_model_mesh = xs.Mesh(device_ids, (data, mod)) | ||
| model_data_mesh = xs.Mesh(device_ids, (mod, data)) |
There was a problem hiding this comment.
Can you try with HybridMesh? It should provide some performance gain, but I haven't actually benchmarked the difference. Here and in modeling_llama.py
@khatwanimohit may have some benchmarked differences on the simple shardings.py script
There was a problem hiding this comment.
Let me do that. Always forgot.
There was a problem hiding this comment.
Fixed that too.
Comment on lines
554
to
557
| elif 'gate_proj' in name or 'up_proj' in name: | ||
| xs.mark_sharding(param, data_model_mesh, range(len(param.shape))) | ||
| elif 'down_proj' in name: | ||
| xs.mark_sharding(param, model_data_mesh, range(len(param.shape))) |
There was a problem hiding this comment.
Just for my understanding: I noticed that HF shards gate_proj and up_proj on the 0th dim and down_proj on the 1st dim, but here you're sharding gate and up on the data_model mesh, which places the model axis on dim 1.
Is this just a difference in 1- and 2-D sharding?
There was a problem hiding this comment.
That's a good catch. I don't know. Let me dig into it. I'm following the slides attached on the top of the spreadsheet.
There was a problem hiding this comment.
No worries! I was just curious, using the sharding from the slides makes sense.
There was a problem hiding this comment.
Yea, you are right. I have corrected the error.
|
Thanks Jon for approving the pull request. |
alanwaketan
added a commit
that referenced
this pull request
Aug 2, 2023
Summary: This pull requests fix a bug in #17 where it forgot to guard 2D sharding for activations and inputs. Test Plan: N/A.
alanwaketan
added a commit
that referenced
this pull request
Oct 27, 2023
Summary: This pull request adds 2D SPMD sharding to the table. It will shard both weights and activations. Here is the sharding strategy. Let's say we have a 2D mesh (data, model) and data x model == num_devices: 1. input (data,, None, model) 2. embedding (model, data) 3. attn QKV (data, model) 4. attn O (model, data) 5. mlp gate, up (model, data) 6. mlp down (data, model) 7. activation (data,, None, model) Currently you can specify the model dimension using a new option --spmd_2d_sharding, then the data dimension will be auto-calculated. TODO: maybe we should have another option to specify whether or not we should shard the activations/inputs or shard them differently.
alanwaketan
added a commit
that referenced
this pull request
Oct 27, 2023
Summary: This pull requests fix a bug in #17 where it forgot to guard 2D sharding for activations and inputs. Test Plan: N/A.
yeounoh
pushed a commit
that referenced
this pull request
Mar 19, 2024
Summary: This pull request adds 2D SPMD sharding to the table. It will shard both weights and activations. Here is the sharding strategy. Let's say we have a 2D mesh (data, model) and data x model == num_devices: 1. input (data,, None, model) 2. embedding (model, data) 3. attn QKV (data, model) 4. attn O (model, data) 5. mlp gate, up (model, data) 6. mlp down (data, model) 7. activation (data,, None, model) Currently you can specify the model dimension using a new option --spmd_2d_sharding, then the data dimension will be auto-calculated. TODO: maybe we should have another option to specify whether or not we should shard the activations/inputs or shard them differently.
yeounoh
pushed a commit
that referenced
this pull request
Mar 19, 2024
Summary: This pull requests fix a bug in #17 where it forgot to guard 2D sharding for activations and inputs. Test Plan: N/A.
vanbasten23
pushed a commit
that referenced
this pull request
May 21, 2024
Summary: This pull request adds 2D SPMD sharding to the table. It will shard both weights and activations. Here is the sharding strategy. Let's say we have a 2D mesh (data, model) and data x model == num_devices: 1. input (data,, None, model) 2. embedding (model, data) 3. attn QKV (data, model) 4. attn O (model, data) 5. mlp gate, up (model, data) 6. mlp down (data, model) 7. activation (data,, None, model) Currently you can specify the model dimension using a new option --spmd_2d_sharding, then the data dimension will be auto-calculated. TODO: maybe we should have another option to specify whether or not we should shard the activations/inputs or shard them differently.
vanbasten23
pushed a commit
that referenced
this pull request
May 21, 2024
Summary: This pull requests fix a bug in #17 where it forgot to guard 2D sharding for activations and inputs. Test Plan: N/A.
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.
Summary:
This pull request adds 2D SPMD sharding to the table. It will shard both weights and activations. Here is the sharding strategy.
Let's say we have a 2D mesh (data, model) and data x model == num_devices:
Currently you can specify the model dimension using a new option --spmd_2d_sharding, then the data dimension will be auto-calculated.
TODO: maybe we should have another option to specify whether or not we should shard the activations/inputs or shard them differently.