EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Eagle3 mm support, enablement on qwen3vl (#4848)

Browse Source 
### What this PR does / why we need it?
follow pr
[https://github.com/vllm-project/vllm/pull/20788](https://github.com/vllm-project/vllm/pull/20788)
, Eagle3 mm support, enablement on qwen3vl
target model
[Qwen/Qwen3-VL-8B-Instruct]([https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct])
eagle3
[MNN/Qwen3-VL-8B-Instruct-Eagle3](https://www.modelscope.cn/models/MNN/Qwen3-VL-8B-Instruct-Eagle3)
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

pytest ./tests/e2e/singlecard/test_completion_with_prompt_embeds.py -vv

vLLM with eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images  --speculative-config '{
    "method": "eagle3",
    "model": "/model/hf/Qwen3-VL-8B-Instruct-Eagle3",
    "num_speculative_tokens": 3
  }'
```
vLLM without eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images 
```

bench:
```
vllm bench serve   --backend openai-chat   --base-url http://127.0.0.1:9100   --tokenizer /model/Qwen3-VL-8B-Instruct   --endpoint /v1/chat/completions   --model /model/Qwen3-VL-8B-Instruct   --dataset-name random  --num-prompts 50   --max-concurrency 5   --temperature 0   --top-p 1.0   --seed 123
```

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: jesse <szxfml@gmail.com>
This commit is contained in:
Song Zhixin
2026-01-19 08:58:07 +08:00
committed by GitHub
parent 05e69b99e5
commit 2b6dc100b5
5 changed files with 145 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.github/workflows/misc/model_list.json vendored
View File

@@ -148,6 +148,7 @@
"moonshotai/Kimi-K2-Thinking",
"moonshotai/Kimi-Linear-48B-A3B-Instruct",
"neuralmagic/Qwen2.5-3B-quantized.w8a8",
"MNN/Qwen3-VL-8B-Instruct-Eagle3",
"nv-community/audio-flamingo-3",
"nv-community/audio-flamingo-3-hf",
"nvidia/audio-flamingo-3-hf",
@@ -234,4 +235,3 @@
"xlangai/OpenCUA-7B"
]
}

50
tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py
View File

@@ -85,6 +85,14 @@ def eagle3_model_name():
return "vllm-ascend/EAGLE3-LLaMA3.1-Instruct-8B"
@pytest.fixture
def vl_model_name():
return "Qwen/Qwen3-VL-8B-Instruct"
def vl_eagle3_model_name():
return "MNN/Qwen3-VL-8B-Instruct-Eagle3"
def test_ngram_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,
@@ -129,6 +137,48 @@ def test_ngram_correctness(
assert matches > int(0.66 * len(ref_outputs))
def test_qwen3_vl_eagle_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,
vl_model_name: str,
):
'''
Compare the outputs of a original LLM and a speculative LLM
should be the same when using eagle speculative decoding.
'''
with VllmRunner(
vl_model_name,
max_model_len=1024,
cudagraph_capture_sizes=[1, 2, 4, 8],
) as ref_llm:
ref_outputs = ref_llm.model.chat(test_prompts, sampling_config)
spec_model_name = vl_eagle3_model_name()
with VllmRunner(
vl_model_name,
speculative_config={
"method": "eagle3",
"model": spec_model_name,
"num_speculative_tokens": 2,
},
max_model_len=1024,
cudagraph_capture_sizes=[1, 2, 4, 8],
) as runner:
spec_outputs = runner.model.chat(test_prompts, sampling_config)
matches = 0
misses = 0
for ref_output, spec_output in zip(ref_outputs, spec_outputs):
if ref_output.outputs[0].text == spec_output.outputs[0].text:
matches += 1
else:
misses += 1
print(f"ref_output: {ref_output.outputs[0].text}")
print(f"spec_output: {spec_output.outputs[0].text}")
# Heuristic: expect at least 70% of the prompts to match exactly
# Upon failure, inspect the outputs to check for inaccuracy.
assert matches > int(0.66 * len(ref_outputs))
def test_suffix_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,

4
tests/ut/spec_decode/test_eagle_proposer.py
View File

@@ -50,6 +50,7 @@ class TestEagleProposerInitialization(TestBase):
self.vllm_config.speculative_config.draft_model_config.get_hidden_size.return_value = 4096
self.vllm_config.compilation_config.mode = CompilationMode.VLLM_COMPILE
self.vllm_config.model_config.enforce_eager = False
self.vllm_config.model_config.uses_mrope = False
self.vllm_config.speculative_config.enforce_eager = False
self.vllm_config.scheduler_config.async_scheduling = False
init_ascend_config(self.vllm_config)
@@ -156,6 +157,7 @@ class TestEagleProposerLoadModel(TestBase):
}]
mock_model = MagicMock()
mock_model.supports_multimodal = False
mock_model.model.embed_tokens = MagicMock()
mock_model.lm_head = MagicMock()
mock_model.multimodal_cpu_fields = None
@@ -226,7 +228,7 @@ class TestEagleProposerLoadModel(TestBase):
self.proposer.name = SpecDcodeType.EAGLE
self.proposer.load_model(mock_model)
mock_model.get_language_model.assert_called_once()
self.assertEqual(mock_model.get_language_model.call_count, 2)
self.assertIs(self.proposer.model.lm_head,
mock_model.get_language_model.return_value.lm_head)

110
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -106,7 +106,7 @@ class EagleProposer(VllmEagleProposer):
self.dcp_size = self.runner.dcp_size
self.pcp_rank = self.runner.pcp_rank
self.dcp_rank = self.runner.dcp_rank
self.use_aclgraph = self.runner._use_aclgraph()
self.full_indices = range(
@@ -149,8 +149,34 @@ class EagleProposer(VllmEagleProposer):
assert len(draft_attn_layer_names) == 1
self.attn_layer_names = list(draft_attn_layer_names)
if supports_multimodal(model):
# handle multimodality
if self.get_model_name(model) in [
"Qwen2_5_VLForConditionalGeneration",
"Qwen3VLForConditionalGeneration",
]:
self.model.config.image_token_index = model.config.image_token_id
elif self.get_model_name(
model) == "PixtralForConditionalGeneration":
self.model.config.image_token_index = (
model.config.vision_config.image_token_id)
else:
self.model.config.image_token_index = (
model.config.image_token_index)
target_language_model = model.get_language_model()
else:
target_language_model = model
# share embed_tokens with the target model if needed
if get_pp_group().world_size == 1:
if hasattr(target_language_model.model, "embed_tokens"):
target_embed_tokens = target_language_model.model.embed_tokens
elif hasattr(target_language_model.model, "embedding"):
target_embed_tokens = target_language_model.model.embedding
else:
raise AttributeError(
"Target model does not have 'embed_tokens' or 'embedding' attribute"
)
if self.method == "mtp":
if self.vllm_config.model_config.is_deepseek_mla and \
torch.equal(self.model.model.embed_tokens.weight,
@@ -161,7 +187,7 @@ class EagleProposer(VllmEagleProposer):
"The MTP head shares the same vocab embedding" \
" with the target model."
)
self.model.model.embed_tokens = model.model.embed_tokens
self.model.model.embed_tokens = target_embed_tokens
else:
logger.info(
" The MTP head loaded its own vocab embedding" \
@@ -172,13 +198,12 @@ class EagleProposer(VllmEagleProposer):
"The EAGLE head shares the same vocab embedding" \
" with the target model."
)
self.model.model.embed_tokens = model.model.embed_tokens
self.model.model.embed_tokens = target_embed_tokens
else:
logger.info(
"Since PP > 1 or other reasons the model head loaded its own vocab embedding" \
" weights instead of sharing them with the target model."
)
# share lm_head with the target model if needed
# some model definition do not define lm_head explicitly
# and reuse embed_tokens for lm_head, e.g., CohereForCausalLM
@@ -221,7 +246,7 @@ class EagleProposer(VllmEagleProposer):
dummy_compute_logits=lambda hidden_states: None,
is_profile=False):
# update global cos, sin
update_cos_sin(self.positions[:num_tokens])
update_cos_sin(self._get_positions(num_tokens))
attn_metadata = None
if not self.use_cuda_graph:
@@ -265,7 +290,7 @@ class EagleProposer(VllmEagleProposer):
attn_metadata[layer_name] = attn_metadata_eagle
model_input_ids = self.input_ids[:num_tokens]
model_positions = self.positions[:num_tokens]
model_positions = self._get_positions(num_tokens)
model_previous_hidden_states = self.hidden_states[:num_tokens]
for i in range(self.num_speculative_tokens):
if i > 0 and in_graph_capturing and aclgraph_runtime_mode == CUDAGraphMode.FULL:
@@ -340,7 +365,6 @@ class EagleProposer(VllmEagleProposer):
# Replace the last token with the next token.
# E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
self.input_ids[last_token_indices] = next_token_ids
if self.use_cuda_graph and \
num_tokens <= self.runner.cudagraph_batch_sizes[-1]:
num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
@@ -356,15 +380,28 @@ class EagleProposer(VllmEagleProposer):
batch_descriptor = None
# copy inputs to buffer for cudagraph
self.positions[:num_tokens] = target_positions
self._set_positions(num_tokens, target_positions)
self.hidden_states[:num_tokens] = target_hidden_states
if self.supports_mm_inputs:
mm_embeds, is_mm_embed = mm_embed_inputs or (None, None)
inputs_embeds = self.model.embed_input_ids(
self.input_ids[:num_tokens],
multimodal_embeddings=mm_embeds,
is_multimodal=is_mm_embed)
self.inputs_embeds[:num_tokens] = inputs_embeds
inputs_embeds = self.inputs_embeds[:num_input_tokens]
input_ids = self.input_ids[:num_input_tokens]
else:
inputs_embeds = None
input_ids = self.input_ids[:num_input_tokens]
# FIXME(woosuk): The below two ops cause synchronization. Optimize.
builder = self.runner.attn_groups[0][0].get_metadata_builder()
attn_metadata = builder.build(0, common_attn_metadata,
self.runner.get_model())
# update global cos, sin
update_cos_sin(self.positions[:num_input_tokens])
update_cos_sin(self._get_positions(num_input_tokens))
per_layer_attn_metadata = {}
for layer_name in self.attn_layer_names:
per_layer_attn_metadata[layer_name] = attn_metadata
@@ -380,7 +417,7 @@ class EagleProposer(VllmEagleProposer):
# The lifecycle of `input_ids`, `positions`, `hidden_states` runs through all speculative tokens' proposings.
# `model_input_ids`, `model_positions` and `model_hidden_states` are used to represent the inputs of speculative model.
model_input_ids = self.input_ids[:num_input_tokens]
model_positions = self.positions[:num_input_tokens]
model_positions = self._get_positions(num_input_tokens)
model_hidden_states = self.hidden_states[:num_input_tokens]
model_hidden_states, model_positions = self.maybe_pad_and_reduce(
@@ -390,6 +427,7 @@ class EagleProposer(VllmEagleProposer):
input_ids=model_input_ids,
positions=model_positions,
hidden_states=model_hidden_states,
inputs_embeds = inputs_embeds
)
if self.method == "mtp":
last_hidden_states = ret_hidden_states
@@ -420,8 +458,10 @@ class EagleProposer(VllmEagleProposer):
dtype=draft_token_ids.dtype,
device=self.device)
draft_token_ids_tensor[0] = draft_token_ids
positions = target_positions[last_token_indices]
if self.uses_mrope:
positions = target_positions[:, last_token_indices]
else:
positions = target_positions[last_token_indices]
hidden_states = hidden_states[last_token_indices]
last_token_indices = self.arange[:batch_size]
@@ -460,11 +500,18 @@ class EagleProposer(VllmEagleProposer):
# but adjust the position ids and slot mappings to avoid the
# out-of-range access during the model execution. The draft tokens
# generated with this adjustment should be ignored.
exceeds_max_model_len = positions >= self.vllm_config.model_config.max_model_len
# Mask out the position ids that exceed the max model length.
# Otherwise, we may get out-of-range error in RoPE.
clamped_positions = torch.where(exceeds_max_model_len, 0,
positions)
if self.uses_mrope:
exceeds_max_model_len = positions[
0] >= self.vllm_config.model_config.max_model_len
# Mask out the position ids that exceed the max model length.
# Otherwise, we may get out-of-range error in RoPE.
clamped_positions = torch.where(
exceeds_max_model_len.unsqueeze(0),
torch.zeros_like(positions), positions)
else:
exceeds_max_model_len = positions >= self.vllm_config.model_config.max_model_len
clamped_positions = torch.where(exceeds_max_model_len, 0,
positions)
# TODO: Increment the sequence lengths.
@@ -485,12 +532,19 @@ class EagleProposer(VllmEagleProposer):
block_size = attn_metadata_builder.kv_cache_spec.block_size
# Compute the slot mapping.
block_numbers = (clamped_positions // block_size)
if self.uses_mrope:
block_numbers = clamped_positions[0] // block_size
else:
block_numbers = (clamped_positions // block_size)
block_ids = attn_metadata.block_tables.gather(
dim=1, index=block_numbers.view(-1, 1))
block_ids = block_ids.view(-1)
slot_mapping_tmp = (block_ids * block_size +
clamped_positions % block_size)
if self.uses_mrope:
slot_mapping_tmp = (block_ids * block_size +
clamped_positions[0] % block_size)
else:
slot_mapping_tmp = (block_ids * block_size +
clamped_positions % block_size)
# Mask out the slot mappings that exceed the max model length.
# Otherwise, the KV cache will be inadvertently updated with the
@@ -504,14 +558,23 @@ class EagleProposer(VllmEagleProposer):
PADDING_SLOT_ID)
# copy inputs to buffer for cudagraph
self.input_ids[:batch_size] = input_ids
self.positions[:batch_size] = clamped_positions
self._set_positions(batch_size, clamped_positions)
self.hidden_states[:batch_size] = hidden_states
if self.supports_mm_inputs:
self.inputs_embeds[:batch_size] = self.model.embed_input_ids(
input_ids)
input_ids = self.input_ids[:input_batch_size]
inputs_embeds = self.inputs_embeds[:input_batch_size]
else:
input_ids = self.input_ids[:input_batch_size]
inputs_embeds = None
attn_mask = self.attn_mask_builder.get_splitfuse_attn_mask()
attn_metadata.attn_mask = attn_mask
# update global cos, sin
update_cos_sin(self.positions[:input_batch_size])
update_cos_sin(self._get_positions(input_batch_size))
# Run the model.
with set_ascend_forward_context(
@@ -526,7 +589,7 @@ class EagleProposer(VllmEagleProposer):
# The lifecycle of `input_ids`, `positions`, `hidden_states` runs through all speculative tokens' proposings.
# `model_input_ids`, `model_positions` and `model_hidden_states` are used to represent the inputs of speculative model.
model_input_ids = self.input_ids[:input_batch_size]
model_positions = self.positions[:input_batch_size]
model_positions = self._get_positions(input_batch_size)
model_hidden_states = self.hidden_states[:input_batch_size]
model_hidden_states, model_positions = self.maybe_pad_and_reduce(
@@ -536,6 +599,7 @@ class EagleProposer(VllmEagleProposer):
input_ids=model_input_ids,
positions=model_positions,
hidden_states=model_hidden_states,
inputs_embeds = inputs_embeds
)
if self.method == "mtp":
last_hidden_states = ret_hidden_states

8
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1354,14 +1354,14 @@ class NPUModelRunner(GPUModelRunner):
query_start_loc_pcp_full[1:num_reqs + 1] - 1
target_token_ids = input_ids_pcp_full[:
num_scheduled_tokens]
target_positions = positions[:num_scheduled_tokens]
target_positions = self._get_positions(num_scheduled_tokens)
target_hidden_states = hidden_states
else:
token_indices_to_sample = None
# input_ids can be None for multimodal models.
target_token_ids = self.input_ids.gpu[:
num_scheduled_tokens]
target_positions = positions[:num_scheduled_tokens]
target_positions = self._get_positions(num_scheduled_tokens)
if self.use_aux_hidden_state_outputs:
target_hidden_states = torch.cat([
h[:num_scheduled_tokens]
@@ -1402,7 +1402,7 @@ class NPUModelRunner(GPUModelRunner):
target_hidden_states = hidden_states
else:
target_token_ids = self.input_ids.gpu[token_indices]
target_positions = positions[token_indices]
target_positions = self._get_positions(token_indices)
if self.use_aux_hidden_state_outputs:
target_hidden_states = torch.cat(
[h[token_indices] for h in aux_hidden_states],
@@ -3006,7 +3006,7 @@ class NPUModelRunner(GPUModelRunner):
def _prepare_multimodal_fields(self):
"""
Ensures specific multimodal tensors are on CPU.
This is necessary for fields like 'grid_thw' which are converted to numpy
This is necessary for fields like 'grid_thw' which are converted to numpy
inside the model's forward pass.
"""
if not self.multimodal_cpu_fields: