merge dev分支代码

vllm/model_executor/models/falcon.py

@@ -58,6 +58,7 @@ from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.utils import pad_weight, gemm_bank_conf
+import vllm.envs as envs
 
 FalconConfig = Union[HF_FalconConfig, RWConfig]
 
@@ -393,7 +394,7 @@ class FalconModel(nn.Module):
         self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
         self.use_fa_pad = os.environ.get('FA_PAD') == '1'
         self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
-        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
+        self.w8a8_strategy = envs.VLLM_W8A8_BACKEND
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.word_embeddings(input_ids)

vllm/model_executor/models/glm4.py

@@ -31,6 +31,7 @@ from typing import Optional, Union
 import torch
 from torch import nn
 from transformers import Glm4Config
+import vllm.envs as envs
 
 
 class MultiModalConfigProxy:
@@ -332,7 +333,7 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
         self.use_fa_pad = os.environ.get('FA_PAD') == '1'
         self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
-        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
+        self.w8a8_strategy = envs.VLLM_W8A8_BACKEND
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.model.get_input_embeddings(input_ids)

vllm/model_executor/models/qwen3.py

@@ -52,6 +52,7 @@ from .qwen2 import Qwen2MLP as Qwen3MLP
 from .qwen2 import Qwen2Model
 from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
 import vllm.envs as envs
+from vllm.utils import direct_register_custom_op
 
 logger = init_logger(__name__)
 
@@ -129,6 +130,58 @@ class Qwen3Attention(nn.Module):
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
 
+    def rms_rotary_embedding_fuse(
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        head_size: int,
+        cos_sin_cache: torch.Tensor,
+        is_neox_style: bool,
+        q_weight: torch.Tensor,
+        k_weight: torch.Tensor,
+        q_bias: Optional[torch.Tensor],
+        k_bias: Optional[torch.Tensor],
+        epsilon: float,
+    ) -> None:
+        from lightop import rms_rotary_embedding_fuse as fused_kernel
+        fused_kernel(
+            positions,
+            query,
+            key,
+            head_size,
+            cos_sin_cache,
+            is_neox_style,
+            q_weight,
+            k_weight,
+            q_bias,
+            k_bias,
+            epsilon,
+        )
+
+    def rms_rotary_embedding_fuse_fake(
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        head_size: int,
+        cos_sin_cache: torch.Tensor,
+        is_neox_style: bool,
+        q_weight: torch.Tensor,
+        k_weight: torch.Tensor,
+        q_bias: Optional[torch.Tensor],
+        k_bias: Optional[torch.Tensor],
+        epsilon: float,
+    ) -> None:
+        # Fake impl intentionally left as no-op for graph tracing modes.
+        pass
+
+    if not hasattr(torch.ops.vllm, "rms_rotary_embedding_fuse"):
+        direct_register_custom_op(
+            op_name="rms_rotary_embedding_fuse",
+            op_func=rms_rotary_embedding_fuse,
+            mutates_args=["query", "key"],
+            fake_impl=rms_rotary_embedding_fuse_fake,
+        )
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -136,7 +189,34 @@ class Qwen3Attention(nn.Module):
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
-        # Add qk-norm
+        if envs.VLLM_USE_FUSED_RMS_ROPE:
+            # Fused RMSNorm + RoPE path through custom op.
+            cos_sin_cache = self.rotary_emb.cos_sin_cache
+            if (cos_sin_cache.device != q.device
+                    or cos_sin_cache.dtype != q.dtype):
+                cos_sin_cache = cos_sin_cache.to(q.device,
+                                                 dtype=q.dtype,
+                                                 non_blocking=True)
+                # Persist the converted cache so we don't re-copy/re-allocate
+                # on every forward when the original buffer starts on CPU.
+                self.rotary_emb.cos_sin_cache = cos_sin_cache
+            q = q.contiguous()
+            k = k.contiguous()
+            torch.ops.vllm.rms_rotary_embedding_fuse(
+                positions,
+                q,
+                k,
+                self.head_dim,
+                cos_sin_cache,
+                self.rotary_emb.is_neox_style,
+                self.q_norm.weight,
+                self.k_norm.weight,
+                None,
+                None,
+                self.q_norm.variance_epsilon,
+            )
+        else:
+            # Add qk-norm then RoPE (original path).
             q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
                                self.head_dim)
             if envs.VLLM_USE_APEX_RN:

vllm/model_executor/models/qwen3_moe.py

@@ -38,6 +38,19 @@ from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe import FusedMoE
+try:
+    from vllm.model_executor.layers.fused_moe.router_capture import (
+        maybe_record_router_logits,
+    )
+except ImportError:
+
+    def maybe_record_router_logits(
+        *,
+        layer_name: str,
+        router_logits: torch.Tensor,
+        top_k: int,
+    ) -> None:
+        return None
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                                QKVParallelLinear,
@@ -111,6 +124,8 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
     ):
         super().__init__()
         self.tp_size = get_tensor_model_parallel_world_size()
+        self._router_top_k = int(config.num_experts_per_tok)
+        self._router_capture_layer_name = prefix
 
         if self.tp_size > config.num_experts:
             raise ValueError(
@@ -140,6 +155,14 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
+        if not (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling()):
+            capture_enabled = envs.VLLM_MOE_ROUTER_CAPTURE
+            if capture_enabled:
+                maybe_record_router_logits(
+                    layer_name=self._router_capture_layer_name,
+                    router_logits=router_logits,
+                    top_k=self._router_top_k,
+                )
         final_hidden_states = self.experts(hidden_states=hidden_states,
                                            router_logits=router_logits)
 
@@ -453,7 +476,7 @@ class Qwen3MoeModel(nn.Module):
         self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
         self.use_fa_pad = os.environ.get('FA_PAD') == '1'
         self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
-        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
+        self.w8a8_strategy = envs.VLLM_W8A8_BACKEND
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.embed_tokens(input_ids)

vllm/model_executor/models/telechat2.py

@@ -37,6 +37,7 @@ from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
 from vllm.utils import W8a8GetCacheJSON
 from vllm import _custom_ops as ops
 from vllm.model_executor.utils import pad_weight, gemm_bank_conf
+import vllm.envs as envs
 
 
 class TeleChat2Model(LlamaModel):
@@ -66,8 +67,7 @@ class TeleChat2Model(LlamaModel):
         self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
         self.use_fa_pad = os.environ.get('FA_PAD') == '1'
         self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
-        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
-
+        self.w8a8_strategy = envs.VLLM_W8A8_BACKEND
     def load_weights(self, weights: Iterable[tuple[str,
                                                    torch.Tensor]]) -> set[str]:
         stacked_params_mapping = [

vllm/utils/__init__.py

@@ -1958,7 +1958,7 @@ class W8a8GetCacheJSON:
         self.moe_weight_shapes=[]
         arch_name = torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0]
         arch_cu = torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count
-        
+        self.cache_json_data = {}
         device_name =arch_name+'_'+str(arch_cu)+'cu'
         self.device_name=device_name
         self.topk=1
@@ -2060,19 +2060,27 @@ class W8a8GetCacheJSON:
         return self.triton_json_dir+f"/linear_{n}_{k}_block[{block_n},{block_k}]_{self.device_name}.json"
 
     def get_moeint8json_name(self,E,N1,N2,K,TOPK,
-                             block_size:Optional[list]=None,use_int4_w4a8:Optional[bool]=False):
+                             block_size:Optional[list]=None,use_int4_w4a8:Optional[bool]=False,use_int8_w8a8:Optional[bool]=False):
         if use_int4_w4a8:
             if block_size is not None:
                 return self.triton_json_dir+f"/MOE_W4A8INT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
             else:
                 return self.triton_json_dir+f"/MOE_W4A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
+        elif use_int8_w8a8:
+            if block_size is not None:
+                return self.triton_json_dir + f"/MOE_BLOCKINT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
+            else:
+                return self.triton_json_dir + f"/MOE_W8A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
         else:
             if block_size is not None:
-                return self.triton_json_dir+f"/MOE_BLOCKINT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
+                return self.triton_json_dir + f"/MOE_BLOCKFP8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
             else:
-                return self.triton_json_dir+f"/MOE_W8A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
+                return self.triton_json_dir + f"/MOE_W8A8FP8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
 
     def get_moeint8_triton_cache(self,file_path,E,N1,N2,K,TOPK):
+        if file_path in self.cache_json_data:
+            # 直接返回缓存数据，避免重复读取
+            return self.cache_json_data[file_path]
         cache_json_file=file_path
         
         if os.path.exists(file_path):
@@ -2088,7 +2096,7 @@ class W8a8GetCacheJSON:
             for sub_key, sub_value in value.items():
                 configs_key= f"{sub_key}_{key}"   
                 configs_dict[configs_key]=sub_value
-    
+        self.cache_json_data[file_path] = configs_dict
         return configs_dict
    
 # Adapted from: https://stackoverflow.com/a/47212782/5082708

vllm/v1/attention/backends/mla/common.py

@@ -217,7 +217,7 @@ from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
                                               CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 from vllm.v1.worker.block_table import BlockTable
-from lightop import fused_rms_norm_rope_contiguous
+from lightop import fused_rms_norm_rope_contiguous, fuse_rmsnorm_rope_quant_qkv
 
 try:
     from vllm.vllm_flash_attn import flash_attn_varlen_func
@@ -1233,6 +1233,47 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
                         kv_cache_dtype_str = "bf16"
                 else:
                     kv_cache_dtype_str = self.kv_cache_dtype
+                
+                if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and kv_cache_dtype_str=="fp8_e4m3" and envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA:
+                    if has_prefill:
+                        fused_rms_norm_rope_contiguous(
+                            positions[:num_actual_toks, ...],
+                            q,
+                            k_pe.squeeze(1),
+                            k_c_normed, # not normed
+                            key_normed[:num_actual_toks, ...], # normed
+                            weight,
+                            cos_sin_cache,
+                            attn_metadata.slot_mapping.flatten(),
+                            kv_cache,
+                            kv_cache_dtype_str,
+                            1.0,
+                            False,
+                            1e-6,
+                        ) 
+                    else:
+                        q_tensor = torch.randn(q.shape[0], num_local_heads, self.qk_nope_head_dim + self.qk_rope_head_dim, dtype=q.dtype, device=q.device)
+                        q_quant = torch.empty_like(q_tensor, dtype=torch.float8_e4m3fn, device=q.device)
+                        q_scale = torch.empty(q.shape[0], dtype=torch.float32, device=q.device)
+                        fuse_rmsnorm_rope_quant_qkv(
+                            positions[:num_actual_toks, ...],
+                            query_nope,
+                            q,
+                            q_quant,
+                            q_scale,
+                            k_pe.squeeze(1),
+                            k_c_normed, # not normed
+                            key_normed[:num_actual_toks, ...], # normed
+                            weight,
+                            cos_sin_cache,
+                            attn_metadata.slot_mapping.flatten(),
+                            kv_cache,
+                            kv_cache_dtype_str,
+                            1.0,
+                            False,
+                            1e-6,
+                        ) 
+                else:
                     fused_rms_norm_rope_contiguous(
                         positions[:num_actual_toks, ...],
                         q,
@@ -1259,12 +1300,15 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
 
         if has_decode:
             assert attn_metadata.decode is not None
+            if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and kv_cache_dtype_str=="fp8_e4m3" and envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA:
+                decode_q = q_quant[:num_decode_tokens]
             decode_q_nope, decode_q_pe = decode_q.split(
                 [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
             # Convert from (B, N, P) to (N, B, P)
             decode_q_nope = decode_q_nope.transpose(0, 1)
             # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
-            decode_ql_nope = torch.bmm(decode_q_nope, self.W_UK_T)
+            # todo: bmm support
+            decode_ql_nope = torch.bmm(q_scale, decode_q_nope, self.W_UK_T) if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and kv_cache_dtype_str=="fp8_e4m3" and envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA else torch.bmm(decode_q_nope, self.W_UK_T) 
             # Convert from (N, B, L) to (B, N, L)
             decode_ql_nope = decode_ql_nope.transpose(0, 1)
 

vllm/v1/attention/backends/mla/flashmla.py

@@ -12,6 +12,7 @@ from vllm.attention.ops.flashmla import (flash_mla_with_kvcache,
                                          flash_mla_with_kvcache_q_nope_pe,
                                          get_mla_metadata,
                                          flash_mla_with_kvcache_fp8,
+                                         flash_mla_with_kvcache_fp8_with_cat,
                                          get_mla_decoding_metadata_dense_fp8,
                                          is_flashmla_supported)
 from vllm.logger import init_logger
@@ -181,6 +182,23 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
         assert attn_metadata.decode is not None
 
         if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and kv_cache_dtype == "fp8_e4m3" and envs.VLLM_USE_FLASH_MLA_FP8:
+            if envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA:
+                o, _ = flash_mla_with_kvcache_fp8_with_cat(
+                    q_nope=q_nope.unsqueeze(1),
+                    q_pe=q_pe.unsqueeze(1),
+                    k_cache=kv_c_and_k_pe_cache.unsqueeze(-2),  # Add head dim of 1
+                    block_table=attn_metadata.decode.block_table,
+                    cache_seqlens=attn_metadata.decode.seq_lens,
+                    head_dim_v=self.kv_lora_rank,
+                    tile_scheduler_metadata=attn_metadata.decode.
+                    tile_scheduler_metadata,
+                    num_splits=attn_metadata.decode.num_splits,
+                    softmax_scale=self.scale,
+                    causal=True,
+                    descale_q = q_scale,
+                    descale_k = k_scale,            
+                )
+            else:
                 if envs.VLLM_USE_OPT_CAT:
                     if q_nope.shape[0] < 1024:
                         from vllm.v1.attention.backends.mla.test_concat import concat_helper_decode

vllm/v1/core/sched/scheduler.py

@@ -1051,11 +1051,12 @@ class Scheduler(SchedulerInterface):
     
     def schedule(self) -> SchedulerOutput:
         if envs.VLLM_USE_PD_SPLIT: 
+            if self.use_mla:
+                if self.full_cuda_graph and self.num_spec_tokens > 0:
                     return self.schedule_split_pd() 
                 else:
-            if self.connector is not None:
                     return self.schedule_default()
-            if self.full_cuda_graph and self.use_mla and self.num_spec_tokens > 0 :
+            else:
                 return self.schedule_split_pd()
         else:
             return self.schedule_default()
@@ -1101,13 +1102,14 @@ class Scheduler(SchedulerInterface):
             req_id = req.request_id
             req_ids.append(req_id)
             num_tokens = req.num_generated_token_ids
+
             if self.use_pp:
                 # When using PP, the scheduler sends the sampled tokens back,
                 # because there's no direct communication between the first-
                 # stage worker and the last-stage worker. Otherwise, we don't
                 # need to send the sampled tokens back because the model runner
                 # will cache them.
-                token_ids = req.all_token_ids[-num_tokens:]
+                token_ids = req.all_token_ids[-num_tokens:] if num_tokens > 0 else []
                 new_token_ids.append(token_ids)
             new_block_ids.append(req_to_new_block_ids[req_id])
             num_computed_tokens.append(req.num_computed_tokens)
@@ -1241,7 +1243,7 @@ class Scheduler(SchedulerInterface):
 
             scheduled_spec_token_ids = (
                 scheduler_output.scheduled_spec_decode_tokens.get(req_id))
-            request.num_generated_token_ids = 1
+            request.num_generated_token_ids = len(generated_token_ids)
             if scheduled_spec_token_ids:
                 # num_computed_tokens represents the number of tokens
                 # processed in the current step, considering scheduled
@@ -1253,7 +1255,6 @@ class Scheduler(SchedulerInterface):
                 num_tokens_rejected = (len(scheduled_spec_token_ids) + 1 -
                                        len(generated_token_ids))
                 request.num_computed_tokens -= num_tokens_rejected
-                request.num_generated_token_ids = len(generated_token_ids)
                 spec_decoding_stats = self.make_spec_decoding_stats(
                     spec_decoding_stats,
                     num_draft_tokens=len(scheduled_spec_token_ids),

vllm/v1/worker/gpu_model_runner.py

@@ -31,7 +31,7 @@ from vllm.distributed.parallel_state import (
     prepare_communication_buffer_for_model,
     get_tensor_model_parallel_world_size)
 from vllm.forward_context import (DPMetadata, get_forward_context,
-                                  set_forward_context)
+                                  set_forward_context, set_profilling)
 from vllm.logger import init_logger
 from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
 from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
@@ -514,14 +514,12 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 new_token_ids = req_data.new_token_ids[i]
                 # Add the sampled token(s) from the previous step (if any).
                 # This doesn't include "unverified" tokens like spec tokens.
-                num_new_tokens = (num_computed_tokens + len(new_token_ids) -
-                                  req_state.num_tokens)
+                num_new_tokens = len(new_token_ids)
                 if num_new_tokens == 1:
-                    # Avoid slicing list in most common case.
                     req_state.output_token_ids.append(new_token_ids[-1])
                 elif num_new_tokens > 0:
                     req_state.output_token_ids.extend(
-                        new_token_ids[-num_new_tokens:])
+                        new_token_ids)
             if len(spec_token_ids) > 0:
                 req_state.spec_token_ids = spec_token_ids
 
@@ -541,6 +539,11 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 # The request is not in the persistent batch.
                 # The request was either preempted and resumed later, or was not
                 # scheduled in the previous step and needs to be added again.
+                if not is_last_rank:
+                    req_state = self.requests[req_id]
+                    self.input_batch.add_request(req_state)
+                    req_index = self.input_batch.req_id_to_index.get(req_id)
+                else:
                     req_ids_to_add.append(req_id)
                     continue
 
@@ -554,6 +557,7 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
             if not is_last_rank:
                 # Add new_token_ids to token_ids_cpu.
                 start_token_index = num_computed_tokens
+                if len(new_token_ids) > 0:
                     end_token_index = num_computed_tokens + 1
                     self.input_batch.token_ids_cpu[
                         req_index,
@@ -1598,6 +1602,11 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
             seq_len = (req_state.num_computed_tokens +
                        scheduler_output.num_scheduled_tokens[req_id])
             if seq_len < req_state.num_tokens:
+                # If we have already started decoding, seeing a "partial prefill"
+                # condition is suspicious and can lead to discarding the sampled
+                # token forever (PP stall).
+                if req_state.output_token_ids:
+                    continue
                 # Ignore the sampled token for partial prefills.
                 # Rewind the generator state as if the token was not sampled.
                 # This relies on cuda-specific torch-internal impl details
@@ -1675,11 +1684,9 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 spec_decode_metadata,
                 attn_metadata,
             )
-
         if spec_token_ids is not None:
             for i in discard_sampled_tokens_req_indices:
                 spec_token_ids[i].clear()
-
         # Clear KVConnector state after all KVs are generated.
         if has_kv_transfer_group():
             get_kv_transfer_group().clear_connector_metadata()
@@ -3460,6 +3467,11 @@ class GPUModelRunnerMTP(GPUModelRunnerBase):
             seq_len = (req_state.num_computed_tokens +
                        scheduler_output.num_scheduled_tokens[req_id])
             if seq_len < req_state.num_tokens:
+                # If we have already started decoding, seeing a "partial prefill"
+                # condition is suspicious and can lead to discarding the sampled
+                # token forever (PP stall).
+                if req_state.output_token_ids:
+                    continue
                 # Ignore the sampled token for partial prefills.
                 # Rewind the generator state as if the token was not sampled.
                 # This relies on cuda-specific torch-internal impl details
@@ -3481,7 +3493,6 @@ class GPUModelRunnerMTP(GPUModelRunnerBase):
             hidden_states[:num_scheduled_tokens],
             scheduler_output,
         )
-        #-----------------------------------
         # Get the valid generated tokens.
         sampled_token_ids = sampler_output.sampled_token_ids
         max_gen_len = sampled_token_ids.shape[-1]