Onboard DeepSeek MoE with shared experts #1242
Open
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
RissyRan
requested review from
gobbleturk,
khatwanimohit,
bvandermoon and
vipannalla
as code owners
gagika
approved these changes
Feb 7, 2025
|
|
||
|
|
||
| class DeepSeekMoELayer(nn.Module): | ||
| """DeepSeek-style MoE layer.""" |
There was a problem hiding this comment.
Can you comment what are main differences of DeepSeekMoELayer from regular MoELayer?
| @@ -371,7 +378,11 @@ def permute(self, inputs, gate_logits): | |||
| inputs_shape = inputs.shape | |||
| inputs_2d = jnp.reshape(inputs, (inputs_shape[0] * inputs_shape[1], inputs_shape[2])) | |||
| weights, selected_experts = jax.lax.top_k(gate_logits, self.num_experts_per_tok) | |||
| weights = jax.nn.softmax(weights.astype(jnp.float32), axis=-1).astype(self.dtype) | |||
| if self.config.decoder_block == "deepseek": | |||
There was a problem hiding this comment.
Can you add some comment how/why it's different for deepseek?
also, perhaps you can move it to a function, as it's used in 2 places, e.g.
def _deepseek_scale_weights(self, weights):
"""Scales weights according to DeepSeek's ... ."""
weights /= weights.sum(-1, keepdims=True)
weights *= self.config.routed_scaling_factor
return weights
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Outdated suggestions cannot be applied.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
Onboard DeepSeek MoE with shared experts (functional first - reference from DeepSeek):
models.pyto handle mixed layers, i.e. both dense and moe layersDeepSeekMoeBlockto reuse the dense and moe blocks so we could use either dropping or dropless for future tuningTests
base_num_decoder_layers: 5&num_experts: 16Checklist
Before submitting this PR, please make sure (put X in square brackets):