[Feature] add DFlash Support#7162
[Feature] add DFlash Support#7162chenaoxuan wants to merge 1 commit intovllm-project:releases/v0.13.0from
Conversation
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
Summary of ChangesHello, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request introduces comprehensive support for DFlash, a parallel speculative decoding algorithm, enabling its use within the Ascend-NPU environment. The changes involve adapting core components like the attention mechanism and rotary embeddings, alongside implementing dedicated logic for DFlash's token proposal and dummy run procedures. This integration aims to enhance the efficiency of speculative decoding by leveraging DFlash's diffusion-based approach for generating multiple candidate tokens simultaneously. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds support for DFlash speculative decoding. The changes span across attention layers, rotary embeddings, and the model runner to accommodate the new decoding logic. My review focuses on improving code maintainability by reducing duplication and addressing potential issues like magic numbers and removed assertions. Overall, the changes seem to be in the right direction to support DFlash.
| if not attn_metadata.causal: | ||
| # for dflash | ||
| attn_output, _ = torch_npu.npu_fused_infer_attention_score( | ||
| query=query, | ||
| key=key, | ||
| value=value, | ||
| block_table=block_table, | ||
| input_layout="TND", | ||
| block_size=block_size, | ||
| actual_seq_lengths=attn_metadata.actual_seq_lengths_q, | ||
| actual_seq_lengths_kv=actual_seq_lengths_kv, | ||
| num_key_value_heads=self.num_kv_heads, | ||
| num_heads=self.num_heads, | ||
| scale=self.scale, | ||
| sparse_mode=0, | ||
| ) | ||
| else: | ||
| # Get workspace from cache or calculate it if not present. | ||
| attn_output, _ = torch_npu.npu_fused_infer_attention_score( | ||
| query=query, | ||
| key=key, | ||
| value=value, | ||
| atten_mask=attn_metadata.attn_mask, | ||
| block_table=block_table, | ||
| input_layout="TND", | ||
| block_size=block_size, | ||
| actual_seq_lengths=attn_metadata.actual_seq_lengths_q, | ||
| actual_seq_lengths_kv=actual_seq_lengths_kv, | ||
| num_key_value_heads=self.num_kv_heads, | ||
| num_heads=self.num_heads, | ||
| scale=self.scale, | ||
| sparse_mode=3, | ||
| ) |
There was a problem hiding this comment.
There's significant code duplication between the if and else branches. Most arguments to torch_npu.npu_fused_infer_attention_score are the same. This can be refactored to improve readability and maintainability by creating a common dictionary of arguments.
common_kwargs = {
"query": query,
"key": key,
"value": value,
"block_table": block_table,
"input_layout": "TND",
"block_size": block_size,
"actual_seq_lengths": attn_metadata.actual_seq_lengths_q,
"actual_seq_lengths_kv": actual_seq_lengths_kv,
"num_key_value_heads": self.num_kv_heads,
"num_heads": self.num_heads,
"scale": self.scale,
}
if not attn_metadata.causal:
# for dflash
attn_output, _ = torch_npu.npu_fused_infer_attention_score(
**common_kwargs,
sparse_mode=0,
)
else:
# Get workspace from cache or calculate it if not present.
attn_output, _ = torch_npu.npu_fused_infer_attention_score(
**common_kwargs,
atten_mask=attn_metadata.attn_mask,
sparse_mode=3,
)| if vllm_config.speculative_config.method == "dflash": | ||
| _cos = torch.ones(1, max_num_batched_tokens * 2, 1, rope_dim, dtype=dtype, device=device) | ||
| _sin = torch.zeros(1, max_num_batched_tokens * 2, 1, rope_dim, dtype=dtype, device=device) | ||
| else: | ||
| _cos = torch.ones(1, | ||
| max_num_batched_tokens, | ||
| 1, | ||
| rope_dim, | ||
| dtype=dtype, | ||
| device=device) | ||
| _sin = torch.zeros(1, | ||
| max_num_batched_tokens, | ||
| 1, | ||
| rope_dim, | ||
| dtype=dtype, | ||
| device=device) |
There was a problem hiding this comment.
The if/else block for creating _cos and _sin tensors contains duplicated code. This can be simplified by using a multiplier to determine the size, making the code more concise and easier to maintain.
size_multiplier = 2 if vllm_config.speculative_config and vllm_config.speculative_config.method == "dflash" else 1
shape = (1, max_num_batched_tokens * size_multiplier, 1, rope_dim)
_cos = torch.ones(*shape, dtype=dtype, device=device)
_sin = torch.zeros(*shape, dtype=dtype, device=device)| draft_indexer_layer_names = indexer_layers - target_indexer_layer_names | ||
| draft_attn_layer_names = draft_attn_layer_names - draft_indexer_layer_names | ||
| assert len(draft_attn_layer_names) == 1 | ||
| self.attn_layer_names = list(sorted(draft_attn_layer_names)) |
There was a problem hiding this comment.
The assertion assert len(draft_attn_layer_names) == 1 was removed. While this might be necessary for DFlash, it removes a safety check for other methods like 'eagle' and 'eagle3'. This could lead to unexpected behavior if the draft model for those methods changes. It would be safer to make this check conditional to avoid potential regressions.
if self.method != "dflash":
assert len(draft_attn_layer_names) == 1, (
f"Expected 1 draft attention layer for method '{self.method}', "
f"but found {len(draft_attn_layer_names)}."
)
self.attn_layer_names = list(sorted(draft_attn_layer_names))| total_num_query_tokens = batch_size * num_query_tokens | ||
| num_kv_tokens = num_context_tokens + total_num_query_tokens | ||
|
|
||
| MASK_TOKEN_ID = 151669 |
There was a problem hiding this comment.
The magic number 151669 is used for MASK_TOKEN_ID. This should be defined as a named constant at the top of the file or in a configuration file to improve readability and maintainability.
MASK_TOKEN_ID = 151669 # TODO: Refactor to a named constant| if self.speculative_config.method == "dflash": | ||
| draft_token_ids = self.drafter._dflash_propose( | ||
| target_token_ids=target_token_ids, | ||
| target_positions=target_positions, | ||
| target_hidden_states=target_hidden_states, | ||
| next_token_ids=next_token_ids, | ||
| last_token_indices=token_indices_to_sample, | ||
| common_attn_metadata=common_attn_metadata | ||
| ) | ||
| else: | ||
| draft_token_ids = self.drafter._propose( | ||
| target_token_ids=target_token_ids, | ||
| target_positions=target_positions, | ||
| target_hidden_states=target_hidden_states, | ||
| next_token_ids=next_token_ids, | ||
| last_token_indices=token_indices_to_sample, | ||
| common_attn_metadata=common_attn_metadata, | ||
| sampling_metadata=sampling_metadata, | ||
| req_scheduled_tokens=req_scheduled_tokens, | ||
| long_seq_metadata=long_seq_metadata, | ||
| num_prefill_reqs=num_prefill_reqs, | ||
| num_decode_reqs=num_decode_reqs, | ||
| scheduler_output=scheduler_output, | ||
| num_scheduled_tokens=num_scheduled_tokens, | ||
| ) |
There was a problem hiding this comment.
There is significant code duplication in the if/else block for calling _dflash_propose and _propose. Many arguments are shared. This can be refactored to improve readability and maintainability by using a dictionary for common arguments.
propose_kwargs = {
"target_token_ids": target_token_ids,
"target_positions": target_positions,
"target_hidden_states": target_hidden_states,
"next_token_ids": next_token_ids,
"last_token_indices": token_indices_to_sample,
"common_attn_metadata": common_attn_metadata,
}
if self.speculative_config.method == "dflash":
draft_token_ids = self.drafter._dflash_propose(**propose_kwargs)
else:
propose_kwargs.update({
"sampling_metadata": sampling_metadata,
"req_scheduled_tokens": req_scheduled_tokens,
"long_seq_metadata": long_seq_metadata,
"num_prefill_reqs": num_prefill_reqs,
"num_decode_reqs": num_decode_reqs,
"scheduler_output": scheduler_output,
"num_scheduled_tokens": num_scheduled_tokens,
})
draft_token_ids = self.drafter._propose(**propose_kwargs)
|
Thanks for the PR. It's great. While we hope to support dflash base on vLLM change. Once vllm-project/vllm#36847 is merged, you can implement Ascend one basing on it to reduce code re-write |
Yes, I noticed that this PR has been merged into vllm-main, but vllm-ascend does not yet support the current commit. The cherry-pick version based on this PR has already started development. |
|
@chenaoxuan I am working on enabling dflash on vllm-ascend -- I downloaded the vllm on main, I also downloaded the vllm-ascend from this code Can you provide your thoughts if this is going to work? What will be the blockers? |
You can temporarily use this vllm version: Link |
|
@chenaoxuan I did progress with the latest versions too. Let's collaborate on this so that we can merge efforts. |
I am delighted to collaborate. How can we further communicate? Or through what form? |
|
@chenaoxuan Great, happy to collaborate! We can coordinate here on the RFC for visibility, and if needed we can also chat for quicker discussion. I’m currently working on adapting vLLM with vllm-ascend using the latest versions, and I also ran simulations for vLLM on GPU. What part are you focusing on? If it’s easier to chat, feel free to reach me on Slack/WeLink (or let me know what platform you prefer). |
I tried to contact you via WeLink, but it seems I don't have the necessary permissions, so I couldn't send the message... |
**This PR is inherited from PR-[7162 ](#7162) and supports the latest vllm-ascend main. The old version is closed.** ### Purpose **We first supported DFlash on Ascend-NPU and then maintained it.** > DFlash ("[DFlash: Block Diffusion for Flash Speculative Decoding](https://arxiv.org/abs/2602.06036)") is a parallel speculative decoding algorithm that generates multiple candidate tokens at once through a diffusion process. Main changes: - Corresponds to the official support of vllm merged PR-[36847](vllm-project/vllm#36847). - Add dflash proposer implementation on the basis of SpecDecodeBaseProposer. - Modify the attention backend and add bidirectional attention branch. - Modify model_runner_v1 to support calling the dflash module. ### Quick Start [!Attention!] As of April 10, vllm-ascend is not compatible with vllm that supports DFlash. Therefore, cherry-pick is required: `cd vllm` `git checkout -b new-branch v0.19.0` `git cherry-pick dc14cbf0c06e8a124bdf0c03e8e267feef60887e` [Weights] Use official DFlash [weights](https://huggingface.co/collections/z-lab/dflash). [Config] --speculative-config '{"num_speculative_tokens": 8, "method":"dflash","model":"weight_path","enforce_eager": true}' ### Test Results #### Acceptance rate Verified with Sglang(GPU) and Vllm(GPU) version of Qwen3-8B-DFlash-b16 in GSM8K dataset. T=0 Draft Tokens = 16, Max Tokens = 2048 | Batch Size | Framework | Mean Acceptance Length | |-----|-----|-----| | 4 | SGlang | 6.07 | | 4 | vLLM | 6.08 | | 4 | vLLM-Ascend | 6.05 | | 8 | SGlang | 6.07 | | 8 | vLLM | 6.08 | | 8 | vLLM-Ascend | 6.06 | | 16 | SGlang | 6.08 | | 16 | vLLM | 6.08 | | 16 | vLLM-Ascend | 6.08 | | 32 | SGlang | 6.08 | | 32 | vLLM | 6.09 | | 32 | vLLM-Ascend | 6.08 | #### Performance Qwen3-8B, DP1/TP1, constructing gsm8k dataset to repeat the input length to 3.5K/output length 1.5K, data num 400, batch_size 16, temperature 0 | Method | Graph Mode | Spec Num | Mean Acceptance Length | TOPT(ms) | Output Token Throughput(token/s)| |-----|-----|-----|-----|-----|-----| | Eagle3 | FULL_DECODE_ONLY | 3 | 2.81 | 16.4 | 943.60(baseline) | | Eagle3 | FULL_DECODE_ONLY | 8 | 3.60 | 19.5 | 795.34(↓15.7%) | | Dflash | PIECEWISE | 8 | 5.25 | 12.4 | 1248.93(↑32.4%) | #### Accuracy Qwen3-8B, DP1/TP1, output length 3.5K, data num 300, batch_size 16, temperature 0 | Method | Graph Mode | Spec Num | Dataset | Accuracy(%) | |-----|-----|-----|-----|-----| | Eagle3 | FULL_DECODE_ONLY | 3 | gsm8k | 84.67 | | Dflash | PIECEWISE | 8 | gsm8k | 85.00 | ### Next Plan - Support FULL_DECODE_ONLY - Support Qwen3.5 - The NPU Triton multi-core is faulty. Currently, only use a single core to process all reqs, which needs to be improved. - Operator optimization: now the maximum number of TND layout's for the FIA operator is 16. Therefore, the maximum sepc_num is 15. Although this issue can be bypassed, the performance will be affected. - vLLM main: https://github.com/vllm-project/vllm/commit/v0.19.0 Signed-off-by: chenaoxuan <cax1165@163.com>
3 participants
The current PR has been closed and moved to PR-8118 to support the latest vllm-ascend main.
Purpose
We first supported DFlash on Ascend-NPU and then maintained it.
Main changes:
Quick Start
Test Results
Verified with Sglang(GPU) and Vllm(GPU) version of Qwen3-8B-DFlash-b16 in GSM8K dataset.
T=0 Draft Tokens = 16, Max Tokens = 2048
Next Plan