add multi-stage guide #234
Open
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
jlamypoirier
approved these changes
Apr 17, 2025
| | `2` | Replicated | Sharded | Sharded | Moderate[^1] | | ||
| | `3` | Sharded | Sharded | Sharded | High[^2] | | ||
|
|
||
| [^1]: Communication overhead for ZeRO Stage 2 is similar to Stage 1, except during (depth-first) gradient accumulation when additional all-reduce operations occur. |
There was a problem hiding this comment.
Technically reduce-scatter
|
|
||
| ### Buffers | ||
|
|
||
| When gradients or weights are sharded, Fast-LLM accumulates partial results in shared *buffers* during forward and backward passes, separately for gradients and weights. These buffers reduce communication overhead by batching gradient or weight updates across GPUs or nodes. The options `num_grad_buffers` and `num_weight_buffers` control the number of buffers used for gradients and weights, respectively. |
There was a problem hiding this comment.
Might be useful to state explicitly how this relates to ZeRO stages:
num_layersbuffers: Store all layers in memory, as in traditional (non-ZeRO) DP2Keep weights/gradients one layer at the time, i.e. ZeRO stage 2/3. Second buffer is there for network overlap.
|
|
||
| By default, Fast-LLM assigns one gradient and weight buffer per stage, where the number of stages equals the total number of logical partitions (stages) of the model. This enables overlapping communication (e.g., data transfers between GPUs or nodes) with computation (actual processing done by each GPU or node). Lower values (e.g., 1) reduce this overlap, potentially increasing communication waiting times. | ||
|
|
||
| Increasing `num_grad_buffers` or `num_weight_buffers` provides more room for overlapping communication with compute. This can help in some setups, especially when stages are imbalanced, but generally isn't necessary. Note that this does not reduce total communication; it just shifts when it happens. |
There was a problem hiding this comment.
Missing transition from the last paragraph, this makes it look like we're going higher than num_layers. Reducing (to 1) is also an option to sacrifice network overlap for lower memory usage.
| - **`stages_per_pipeline_stage`**: Intended to specify how many stages run per pipeline worker when pipeline parallelism is active. | ||
|
|
||
| !!! warning | ||
| This feature is currently **not implemented**. Changing this value has no effect. |
There was a problem hiding this comment.
Technically validation will fail
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.
✨ Description
add short multi-stage guide
🔍 Type of change
Select all that apply: