[Hardware][Ascend]forward_oot for FusedMoE #68
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
|
SidaoY
force-pushed
the
v0.7.1-dev
branch
6 times, most recently
from
42d855d to
49a14d2
Compare
SidaoY
changed the title
vllm_ascend/ops/fused_moe.py
Outdated
| num_tokens, _ = hidden_states.shape | ||
| E, N, _ = w1.shape | ||
|
|
||
| batch_size_decode = 1 |
There was a problem hiding this comment.
Why hard code batch_size_decode as 1? We only support bs1 scenario?
vllm_ascend/ops/fused_moe.py
Outdated
| row_idx_decode_len = batch_size_decode * top_k | ||
| row_idx_decode = torch.arange( | ||
| 0, row_idx_decode_len, | ||
| dtype=torch.int32).view(top_k, -1).permute(1, 0).int().contiguous().npu() |
There was a problem hiding this comment.
Seems the same?
Suggested change
| dtype=torch.int32).view(top_k, -1).permute(1, 0).int().contiguous().npu() | |
| torch.arange(0, row_idx_decode_len, dtype=torch.int32, device="npu").view(-1 top_k) |
There was a problem hiding this comment.
Need to confirm if the batch_size_decode can always be hardcode to 1 before adapt this.
vllm_ascend/ops/fused_moe.py
Outdated
| row_idx_prefill_len = batch_size_prefill * num_tokens * top_k | ||
| row_idx = torch.arange( | ||
| 0, row_idx_prefill_len, dtype=torch.int32, | ||
| device=topk_weights.device).view(top_k, -1).permute(1, 0).int().contiguous() |
vllm_ascend/ops/fused_moe.py
Outdated
| expert_tokens = torch_npu.npu_moe_compute_expert_tokens(expanded_expert_idx, E) | ||
| expert_tokens = expert_tokens.to(torch.int64) | ||
|
|
||
| w1 = torch_npu.npu_transpose(w1, (0, 2, 1), require_contiguous=False) |
There was a problem hiding this comment.
Better use torch.transpose? Since require_contiguous=False tranpose() should only change its meta_data, which should be the same with torch.transpose().
SidaoY
force-pushed
the
v0.7.1-dev
branch
4 times, most recently
from
0efd905 to
5ad81ee
Compare
Signed-off-by: YHT <[email protected]>
ganyi1996ppo
approved these changes
Feb 18, 2025
Angazenn
pushed a commit
to Angazenn/vllm-ascend
that referenced
this pull request
Feb 21, 2025
Add npu implement for FusedMoE Signed-off-by: YHT <[email protected]> Co-authored-by: YHT <[email protected]> Signed-off-by: angazenn <[email protected]>
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.
What this PR does / why we need it?
In order to adapt the DeepSeek model of the vLLM framework to Ascend hardware, a Fused MoE (Mixture of Experts) module is developed.
Does this PR introduce any user-facing change?
I've written the NPU version of the group_topk function and the fused_expert function. Next, I'll further implement the forward_oot method of the UnquantizedFusedMoEMethod class.
How was this patch tested?