feat: add mtp ut and fix some bugs (#2453)

### What this PR does / why we need it?
Fix mtp mode ut

### Does this PR introduce _any_ user-facing change?
Nothing

### How was this patch tested?
This can be tested in the same way as a unit test.


- vLLM version: v0.10.0
- vLLM main:
53415653ff

Signed-off-by: 赵江江 <zhaojiangjiang1@h-partners.com>
Co-authored-by: 赵江江 <zhaojiangjiang1@h-partners.com>

tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py

@@ -1,43 +1,13 @@
 from __future__ import annotations
 
-import random
+import os
-from typing import Any
 
 import pytest
-from vllm import LLM, SamplingParams
+from vllm import SamplingParams
 
+from tests.e2e.conftest import VllmRunner
 
-@pytest.fixture
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-def test_prompts():
-    prompt_types = ["repeat", "sentence"]
-    num_prompts = 10
-    prompts = []
-
-    random.seed(0)
-    random_prompt_type_choices = random.choices(prompt_types, k=num_prompts)
-
-    # Generate a mixed batch of prompts, some of which can be easily
-    # predicted by n-gram matching and some which likely cannot.
-    for kind in random_prompt_type_choices:
-        word_choices = ["test", "temp", "hello", "where"]
-        word = random.choice(word_choices)
-        if kind == "repeat":
-            prompt = f"""
-            please repeat the word '{word}' 10 times.
-            give no other output than the word at least ten times in a row,
-            in lowercase with spaces between each word and without quotes.
-            """
-        elif kind == "sentence":
-            prompt = f"""
-            please give a ten-word sentence that
-            uses the word {word} at least once.
-            give no other output than that simple sentence without quotes.
-            """
-        else:
-            raise ValueError(f"Unknown prompt type: {kind}")
-        prompts.append([{"role": "user", "content": prompt}])
-
-    return prompts
 
 
 @pytest.fixture
@@ -50,39 +20,56 @@ def model_name():
     return "wemaster/deepseek_mtp_main_random_bf16"
 
 
-@pytest.mark.skipif(
-    True, reason="TODO: Enable me after test_mtp_correctness is fixed")
 def test_mtp_correctness(
-    test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,
     model_name: str,
 ):
+    example_prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
     '''
     Compare the outputs of a original LLM and a speculative LLM
     should be the same when using mtp speculative decoding.
     '''
-    ref_llm = LLM(model=model_name, max_model_len=256, enforce_eager=True)
+    with VllmRunner(model_name,
-    ref_outputs = ref_llm.chat(test_prompts, sampling_config)
+                    tensor_parallel_size=1,
-    del ref_llm
+                    gpu_memory_utilization=0.7,
+                    max_model_len=256,
+                    enforce_eager=True) as ref_llm:
+        ref_outputs = ref_llm.generate(example_prompts, sampling_config)
+
+    with VllmRunner(
+            model_name,
+            tensor_parallel_size=1,
+            max_num_seqs=256,
+            gpu_memory_utilization=0.7,
+            distributed_executor_backend="mp",
+            enable_expert_parallel=True,
+            speculative_config={
+                "method": "deepseek_mtp",
+                "num_speculative_tokens": 1,
+            },
+            enforce_eager=True,
+            max_model_len=2000,
+            additional_config={"ascend_scheduler_config": {
+                "enabled": False
+            }}) as spec_llm:
+        spec_outputs = spec_llm.generate(example_prompts, sampling_config)
 
-    spec_llm = LLM(model=model_name,
-                   trust_remote_code=True,
-                   speculative_config={
-                       "method": "deepseek_mtp",
-                       "num_speculative_tokens": 1,
-                   },
-                   max_model_len=256,
-                   enforce_eager=True)
-    spec_outputs = spec_llm.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:
+        ref_token_ids = ref_output[0][0]
+        spec_token_ids = spec_output[0][0]
+        if ref_token_ids == spec_token_ids[:len(ref_token_ids)]:
             matches += 1
         else:
             misses += 1
-            print(f"ref_output: {ref_output.outputs[0].text}")
+            print(f"ref_output: {ref_output[1][0]}")
-            print(f"spec_output: {spec_output.outputs[0].text}")
+            print(f"spec_output: {spec_output[1][0]}")
 
     # Heuristic: expect at least 66% of the prompts to match exactly
     # Upon failure, inspect the outputs to check for inaccuracy.

tests/ut/quantization/test_quant_config.py

@@ -113,6 +113,7 @@ class TestAscendQuantConfig(TestBase):
     def test_get_quant_method_for_fused_moe(self):
         fused_moe_layer = MagicMock(spec=FusedMoE)
         fused_moe_layer.moe = MagicMock(spec=FusedMoEConfig)
+        fused_moe_layer.moe_config = MagicMock(spec=FusedMoEConfig)
 
         # Test skipped layer
         with patch.object(self.ascend_config, 'is_layer_skipped_ascend', return_value=True), \

tests/ut/torchair/test_torchair_mla.py

@@ -1,11 +1,13 @@
 from unittest.mock import MagicMock, patch
 
 import torch
+from torch import nn
 from vllm.distributed.parallel_state import GroupCoordinator
 from vllm.model_executor.layers.linear import LinearBase
 
 from tests.ut.base import TestBase
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
+from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
 from vllm_ascend.torchair.torchair_mla import (
     AscendMLATorchairBackend, AscendMLATorchairDecodeMetadata,
     AscendMLATorchairImpl, AscendMLATorchairMetadata,
@@ -398,6 +400,68 @@ class TestAscendMLATorchairMetadataBuilder(TestBase):
         assert torch.equal(sin_golden, metadata.decode.sin)
         assert torch.equal(cos_golden, metadata.decode.cos)
 
+    @patch("vllm_ascend.torchair.torchair_mla.get_ascend_config")
+    def test_build_decode(self, mock_ascend_config):
+        ascend_config = MagicMock()
+        mock_ascend_config.return_value = ascend_config
+        ascend_config.torchair_graph_config.enabled = False
+
+        mock_vllm_config = MagicMock()
+        mock_vllm_config.model_config.max_model_len = 1024
+        mock_vllm_config.cache_config.block_size = 16
+        mock_vllm_config.scheduler_config.chunked_prefill_enabled = False
+        mock_vllm_config.get_head_size.return_value = 64
+        mock_vllm_config.model_config.dtype = torch.float16
+        mock_device = 'cpu'
+        model = MagicMock(spec=nn.Module)
+        model.model = MagicMock(spec=nn.Module)
+
+        builder = AscendMLATorchairMetadataBuilder(
+            mock_vllm_config,
+            mock_device,
+            metadata_cls=AscendMLATorchairMetadata)
+        builder.rope_dim = 64
+
+        builder.sin_cache = torch.tensor([10, 10])
+        builder.cos_cache = torch.tensor([10, 10])
+
+        with patch.object(builder,
+                          "_get_graph_runner_block_tables",
+                          side_effect=lambda x, y: y):
+            common_attn_metadata = AscendCommonAttentionMetadata(
+                query_start_loc=torch.tensor([0, 1, 2, 3]),
+                query_start_loc_cpu=torch.tensor([0, 1, 2, 3]),
+                seq_lens_cpu=torch.tensor([1, 1, 1]),
+                num_reqs=3,
+                num_actual_tokens=3,
+                max_query_len=1,
+                decode_token_per_req=torch.tensor([1, 1, 1]),
+                block_table_tensor=torch.zeros((10, 10)),
+                slot_mapping_cpu=torch.tensor(range(20)),
+                actual_seq_lengths_q=torch.tensor([0, 1, 2]),
+                positions=torch.tensor([1, 1]),
+                attn_mask=torch.ones((15, 15)),
+                spec_attn_mask=None,
+                attn_state=AscendAttentionState.ChunkedPrefill)
+
+            metadata = builder.build(common_attn_metadata, model)
+
+        self.assertIsInstance(metadata, AscendMLATorchairMetadata)
+        self.assertEqual(metadata.num_input_tokens, 0)
+        self.assertEqual(metadata.num_actual_tokens, 3)
+        self.assertEqual(metadata.num_decodes, 3)
+        self.assertEqual(metadata.num_decode_tokens, 3)
+        self.assertEqual(metadata.num_prefills, 0)
+        self.assertEqual(metadata.attn_state,
+                         AscendAttentionState.ChunkedPrefill)
+        self.assertIsNone(metadata.prefill)
+        self.assertIsInstance(metadata.decode, AscendMLATorchairDecodeMetadata)
+        self.assertEqual(metadata.block_tables.shape[0], 3)
+        self.assertEqual(metadata.block_tables.shape[1], 10)
+        self.assertEqual(metadata.seq_lens.shape[0], 3)
+        self.assertEqual(metadata.slot_mapping.shape[0], 3)
+        self.assertEqual(metadata.query_start_loc.shape[0], 4)
+
 
 class TestAscendMLATorchairImpl(TestBase):
 

vllm_ascend/attention/mla_v1.py

@@ -374,18 +374,12 @@ class AscendMLAMetadataBuilder:
 
         decode_metadata = None
         if num_decodes > 0:
-            actual_seq_lengths_q = query_start_loc[1:].tolist()
+            actual_seq_lengths_q = query_start_loc[1:num_decodes + 1].tolist()
             max_seq_lens = seq_lens[:num_decodes].max().item()
             seq_lens = seq_lens[:num_decode_tokens]
             input_positions = input_positions[:num_decode_tokens]
             block_table = block_table[:num_decode_tokens, ...]
             seq_lens_list = seq_lens.tolist()
-            # TODO(xyx): whether this block is necessary without torchair
-            # mtp torchair + PD scenario, last element of actual_seq_lengths_q must equal to batch_size(num_tokens)
-            batch_size = slot_mapping.size(0)
-            if actual_seq_lengths_q[-1] != batch_size \
-                and common_attn_metadata.attn_state == AscendAttentionState.SpecDecoding:
-                actual_seq_lengths_q[-1] = batch_size
 
             cos = self.cos_cache[input_positions].unsqueeze(  # type: ignore
                 1).unsqueeze(2)

vllm_ascend/ops/fused_moe.py

@@ -1178,7 +1178,7 @@ class AscendFusedMoE(FusedMoE):
         if self.scoring_func != "softmax" and not self.use_grouped_topk:
             raise ValueError("Only softmax scoring function is supported for "
                              "non-grouped topk.")
-        self.moe = FusedMoEConfig.make(
+        moe = FusedMoEConfig.make(
             num_experts=self.global_num_experts,
             experts_per_token=top_k,
             hidden_dim=hidden_size,
@@ -1188,8 +1188,10 @@ class AscendFusedMoE(FusedMoE):
             in_dtype=params_dtype,
             quant_config=quant_config)
 
+        self.moe_config = moe
+
         if quant_config is None:
-            self.quant_method = AscendUnquantizedFusedMoEMethod(self.moe)
+            self.quant_method = AscendUnquantizedFusedMoEMethod(moe)
         else:
             self.quant_method = quant_config.get_quant_method(self, prefix)
 

vllm_ascend/quantization/quant_config.py

@@ -102,7 +102,7 @@ class AscendQuantConfig(QuantizationConfig):
         elif isinstance(layer, FusedMoE):
             if self.is_layer_skipped_ascend(prefix,
                                             self.packed_modules_mapping):
-                return AscendUnquantizedFusedMoEMethod(layer.moe)
+                return AscendUnquantizedFusedMoEMethod(layer.moe_config)
             return AscendFusedMoEMethod(self, prefix,
                                         self.packed_modules_mapping)
         elif isinstance(layer, VocabParallelEmbedding):

vllm_ascend/torchair/torchair_mla.py

@@ -492,17 +492,17 @@ class AscendMLATorchairMetadataBuilder:
         graph_pad_size = common_attn_metadata.graph_pad_size
         use_torchair_graph = graph_pad_size != -1
         if num_decodes > 0:
-            actual_seq_lengths_q = query_start_loc[1:].tolist()
+            actual_seq_lengths_q = query_start_loc[1:num_decodes + 1].tolist()
             max_seq_lens = seq_lens[:num_decodes].max().item()
             seq_lens = seq_lens[:num_decode_tokens]
             input_positions = input_positions[:num_decode_tokens]
             block_table = block_table[:num_decode_tokens, ...]
+            num_token_pad_size = 0
             if use_torchair_graph and common_attn_metadata.attn_state in [
                     AscendAttentionState.DecodeOnly,
                     AscendAttentionState.SpecDecoding
             ]:
                 num_reqs_pad_size = 0
-                num_token_pad_size = 0
                 if graph_pad_size != 0:
                     pad_value = 0
                     num_token_pad_size = graph_pad_size - num_decode_tokens
@@ -535,13 +535,14 @@ class AscendMLATorchairMetadataBuilder:
                                                device=input_positions.device)
                 input_positions = torch.cat(
                     [input_positions, position_padding])
-                actual_seq_lengths_q = query_start_loc[1:].tolist(
+                actual_seq_lengths_q = (
-                ) + common_attn_metadata.actual_seq_lengths_q[
+                    actual_seq_lengths_q + common_attn_metadata.
-                    num_reqs:num_reqs + num_reqs_pad_size]
+                    actual_seq_lengths_q[num_reqs:num_reqs +
+                                         num_reqs_pad_size])
             else:
                 seq_lens_list = seq_lens.tolist()
             # mtp torchair + PD scenario, last element of actual_seq_lengths_q must equal to batch_size(num_tokens)
-            batch_size = slot_mapping.size(0)
+            batch_size = num_decode_tokens + num_token_pad_size
             if actual_seq_lengths_q[-1] != batch_size \
                 and common_attn_metadata.attn_state == AscendAttentionState.SpecDecoding:
                 actual_seq_lengths_q[-1] = batch_size

vllm_ascend/worker/mtp_proposer_v1.py

@@ -190,11 +190,6 @@ class MtpProposer:
         self.positions[:num_tokens] = target_positions
         self.hidden_states[:num_tokens] = target_hidden_states
 
-        if attn_metadata.prefill is not None:
-            attn_metadata.prefill.query_lens = query_lens.cpu()
-            attn_metadata.prefill.input_positions = target_positions
-            attn_metadata.prefill.seq_lens = seq_lens
-
         if not self.torchair_graph_enabled:
             # torch mode need to update num_tokens_across_dp
             # TODO: adapt enable_dbo later
@@ -213,6 +208,7 @@ class MtpProposer:
                 num_tokens=num_input_tokens,
                 with_prefill=with_prefill,
                 num_tokens_across_dp=num_tokens_across_dp,
+                reserved_mc2_mask=self.runner.reserved_mc2_mask,
                 in_profile_run=self.runner.in_profile_run,
                 num_actual_tokens=num_tokens):
             with ProfileExecuteDuration().capture_async('mtp_forward'):
@@ -315,6 +311,7 @@ class MtpProposer:
                 num_tokens=num_tokens,
                 with_prefill=with_prefill,
                 num_tokens_across_dp=num_tokens_across_dp,
+                reserved_mc2_mask=self.runner.reserved_mc2_mask,
                 in_profile_run=self.runner.in_profile_run,
                 num_actual_tokens=0):
             if is_running_torchair:

vllm_ascend/worker/worker_v1.py

@@ -47,9 +47,14 @@ from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
 from vllm_ascend.platform import NPUPlatform
 from vllm_ascend.utils import (init_ascend_soc_version,
                                register_ascend_customop, sleep_mode_enabled,
-                               try_register_lib)
+                               try_register_lib, vllm_version_is)
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
+if not vllm_version_is("0.10.1.1"):
+    from vllm.v1.outputs import DraftTokenIds
+else:
+    DraftTokenIds = None
+
 
 class NPUWorker(WorkerBase):
 
@@ -343,3 +348,6 @@ class NPUWorker(WorkerBase):
 
     def get_supported_tasks(self) -> "tuple[SupportedTask, ...]":
         return self.model_runner.get_supported_tasks()
+
+    def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
+        return self.model_runner.take_draft_token_ids()