Merge remote-tracking branch 'origin/v0.9.2-dev' into v0.9.2-dev

vllm/attention/backends/dual_chunk_flash_attn.py

@@ -19,8 +19,13 @@ from vllm.attention.backends.flash_attn import (FlashAttentionBackend,
 from vllm.distributed.parallel_state import get_tensor_model_parallel_rank
 from vllm.logger import init_logger
 from vllm.utils import async_tensor_h2d
-from vllm.vllm_flash_attn import (flash_attn_varlen_func,
-                                  flash_attn_with_kvcache, sparse_attn_func)
+from vllm.platforms import current_platform
+if not current_platform.is_rocm():
+    from vllm.vllm_flash_attn import (flash_attn_varlen_func,
+                                    flash_attn_with_kvcache, sparse_attn_func)
+else:
+    from flash_attn import (flash_attn_varlen_func,
+                        flash_attn_with_kvcache, sparse_attn_func)
 
 if TYPE_CHECKING:
     from vllm.worker.model_runner import ModelInputForGPUBuilder

vllm/attention/backends/utils.py

@@ -246,12 +246,33 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
                 device, non_blocking=True)
 
         else:
-            block_tables = make_tensor_with_pad(
-                self.block_tables,
-                pad=0,
-                dtype=torch.int,
-                device=device,
-            )
+            has_empty: bool = any(len(bt) == 0 for bt in self.block_tables)
+            has_non_empty = any(len(bt) > 0 for bt in self.block_tables)
+            max_block_length = 0
+            if has_empty and has_non_empty:
+                for inter_data in self.input_builder.inter_data_list:
+                    block_tables = inter_data.block_tables
+                    if block_tables:
+                        for seq_id in inter_data.seq_ids:
+                            if seq_id in block_tables:
+                                block_table = block_tables[seq_id]
+                                max_block_length = max(max_block_length, len(block_table))
+            if max_block_length >0:
+                block_tables = make_tensor_with_pad(
+                    self.block_tables,
+                    pad=0,
+                    dtype=torch.int,
+                    device=device,
+                    max_len=max_block_length,
+                )
+            else:
+                 block_tables = make_tensor_with_pad(
+                    self.block_tables,
+                    pad=0,
+                    dtype=torch.int,
+                    device=device,
+                )
+
         assert max_query_len > 0, "query_lens: {}".format(query_lens)
 
         assert device is not None

vllm/config.py

@@ -893,7 +893,8 @@ class ModelConfig:
         optimized_quantization_methods = [
             "fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
             "awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
-            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas"
+            "quark", "modelopt_fp4", "bitblas", "gptq_bitblas",
+            "slimquant_w4a8","slimquant_w4a8_marlin"
         ]
         if self.quantization is not None:
             self.quantization = cast(me_quant.QuantizationMethods,
@@ -920,6 +921,7 @@ class ModelConfig:
                 "awq_marlin",
                 "ipex",
                 "moe_wna16",
+                "slimquant_w4a8_marlin"
             ]
             quantization_methods = [
                 q for q in supported_quantization if q not in overrides

vllm/engine/arg_utils.py

@@ -1107,8 +1107,8 @@ class EngineArgs:
                 "Cuda graph is not supported with DualChunkFlashAttention. "
                 "To run the model in eager mode, set 'enforce_eager=True' "
                 "or use '--enforce-eager' in the CLI.")
-            assert current_platform.is_cuda(), (
-                "DualChunkFlashAttention is only supported on CUDA platform.")
+            assert current_platform.is_cuda() or current_platform.is_rocm(), (
+                "DualChunkFlashAttention is supported on CUDA/ROCM platform.")
             assert not use_v1, (
                 "DualChunkFlashAttention is not supported on V1 engine. "
                 "To run the model in V0 engine, try set 'VLLM_USE_V1=0'")

vllm/model_executor/layers/fused_moe/layer.py

@@ -811,9 +811,9 @@ class FusedMoE(torch.nn.Module):
                     "CompressedTensorsWNA16MoEMethod")):
             moe_quant_params["intermediate_size_full"] = intermediate_size
 
-        if (self.quant_method.__class__.__name__ in ("BlockInt8MoEMethod")):
-            moe_quant_params["intermediate_size"] = self.intermediate_size_per_partition
-        if (self.quant_method.__class__.__name__ in ("SlimQuantW4A8Int8MoEMethod")):
+        if (self.quant_method.__class__.__name__ in ("BlockInt8MoEMethod",
+                                                     "SlimQuantW4A8Int8MoEMethod",
+                                                     "SlimQuantW4A8Int8MarlinMoEMethod")):
             moe_quant_params["intermediate_size"] = self.intermediate_size_per_partition
 
 

vllm/model_executor/layers/quantization/__init__.py

@@ -37,7 +37,8 @@ QuantizationMethods = Literal[
     "auto-round",
     "rtn",
     "blockwise_int8",
-    "slimquant_w4a8"
+    "slimquant_w4a8",
+    "slimquant_w4a8_marlin"
 ]
 QUANTIZATION_METHODS: list[str] = list(get_args(QuantizationMethods))
 
@@ -118,6 +119,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
     from .tpu_int8 import Int8TpuConfig
     from .blockwise_int8 import BlockInt8Config
     from .slimquant_w4a8 import SlimQuantW4A8Int8Config
+    from .slimquant_w4a8_marlin import SlimQuantW4A8Int8MarlinConfig
 
     method_to_config: dict[str, type[QuantizationConfig]] = {
         "aqlm": AQLMConfig,
@@ -151,6 +153,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
         "rtn": RTNConfig,
         "blockwise_int8": BlockInt8Config,
         "slimquant_w4a8":SlimQuantW4A8Int8Config,
+        "slimquant_w4a8_marlin":SlimQuantW4A8Int8MarlinConfig,
     }
     # Update the `method_to_config` with customized quantization methods.
     method_to_config.update(_CUSTOMIZED_METHOD_TO_QUANT_CONFIG)

vllm/model_executor/layers/quantization/slimquant_w4a8_marlin.py

+from typing import Any, Callable, Dict, List, Optional
+import os
+import torch
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.model_executor.utils import set_weight_attrs
+from vllm.distributed import get_tensor_model_parallel_world_size
+from torch.nn.parameter import Parameter
+from vllm.model_executor.layers.quantization import QuantizationMethods
+from vllm.model_executor.layers.linear import (LinearBase,LinearMethodBase)
+from vllm.model_executor.layers.quantization.base_config import (QuantizationConfig,
+                                                                 QuantizeMethodBase)
+from vllm.model_executor.layers.quantization.utils.w4a8_utils import w4a8_2_marlin_weight
+from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
+                                                  FusedMoeWeightScaleSupported)
+from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
+                                           ModelWeightParameter)
+from vllm.model_executor.layers.quantization.slimquant_w4a8 import SlimQuantW4A8Int8LinearMethod
+
+try:
+    from lmslim.layers.fused_moe.fuse_moe_w4a8_marlin import fused_experts_impl_w4a8_marlin
+except Exception:
+    print("INFO: Please install lmslim if you want to infer the quantitative model of moe.\n")
+
+class MarlinMoeWorkspace:
+    """
+    Singleton manager for device-specific workspace buffers used by w4a8 Marlin-MoE.
+    global_reduce_buffer will take 1.5MB * cus (about 120MB for BW200) memoery in each device
+    """
+    _instances = {}
+    def __new__(cls, device):
+        if device not in cls._instances:
+            instance = super().__new__(cls)
+            instance._initialized = False
+            cls._instances[device] = instance
+        return cls._instances[device]
+
+    def __init__(self, device):
+        if self._initialized:
+            return
+        sms = torch.cuda.get_device_properties(device).multi_processor_count
+        self.workspace = torch.zeros(
+            500, dtype=torch.int, device=device, requires_grad=False
+        )
+        self.global_reduce_buffer = torch.zeros(
+            sms * 6 * 128 * 512, dtype=torch.int, device=device, requires_grad=False
+        )
+        self._initialized = True
+
+    def get_buffers(self):
+        return self.workspace, self.global_reduce_buffer
+
+def baseline_scaled_mm(a: torch.Tensor,
+                      b: torch.Tensor,
+                      scale_a: torch.Tensor,
+                      scale_b: torch.Tensor,
+                      out_dtype: torch.dtype,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+    scales= scale_a* scale_b.T
+    gemmout= torch.mm(
+        a.to(dtype=torch.float32), b.to(dtype=torch.float32))
+    output = (scales *gemmout).to(out_dtype)
+    if bias is not None:
+        output = output + bias
+    return output.to(out_dtype)
+
+
+class SlimQuantW4A8Int8MarlinConfig(QuantizationConfig):
+    """Config class for W4A8 Int8 Quantization.
+    - Weight: static, per-channel, symmetric
+    - Activation: dynamic, per-token, symmetric
+    """
+
+    def __init__(self):
+        pass
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.float16, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 75
+
+    @classmethod
+    def get_name(self) -> str:
+        return "slimquant_w4a8_marlin"
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "SlimQuantW4A8Int8MarlinConfig":
+        return cls()
+    @classmethod
+    def override_quantization_method(
+            cls, hf_quant_cfg, user_quant) -> Optional[QuantizationMethods]:
+        if hf_quant_cfg.get("quant_method") == "slimquant_w4a8" \
+                and user_quant == "slimquant_w4a8_marlin":
+            return cls.get_name()
+        return None
+    def get_quant_method(
+        self,
+        layer: torch.nn.Module,
+        prefix: str,
+    ) -> Optional["QuantizeMethodBase"]:
+
+        if isinstance(layer, LinearBase):
+            return SlimQuantW4A8Int8LinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            return SlimQuantW4A8Int8MarlinMoEMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class SlimQuantW4A8Int8MarlinMoEMethod:
+    """MoE method for W4A8INT8 Marlin.
+    Supports loading INT8 checkpoints with static weight scale and
+    dynamic/static activation scale.
+    Args:
+        quant_config: The quantization config.
+    """
+
+    def __new__(cls, *args, **kwargs):
+
+        if not hasattr(cls, "_initialized"):
+            original_init = cls.__init__
+            new_cls = type(
+                cls.__name__,
+                (FusedMoEMethodBase,),
+                {
+                    "__init__": original_init,
+                    **{k: v for k, v in cls.__dict__.items() if k != "__dict__"},
+                },
+            )
+            obj = super(new_cls, new_cls).__new__(new_cls)
+            obj.__init__(*args, **kwargs)
+            return obj
+        return super().__new__(cls)
+
+    def __init__(self, quant_config):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        tp_size = get_tensor_model_parallel_world_size()
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts, 2 * intermediate_size, hidden_size//2, dtype=torch.int8
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(num_experts, hidden_size, intermediate_size//2, dtype=torch.int8),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        w13_weight_scale = torch.nn.Parameter(
+            torch.ones(num_experts, 2 * intermediate_size, 1, dtype=torch.float32),
+            requires_grad=False,
+        )
+        w2_weight_scale = torch.nn.Parameter(
+            torch.ones(num_experts, hidden_size, 1, dtype=torch.float32),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
+        )
+
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        w13_input_scale = None
+        layer.register_parameter("w13_input_scale", w13_input_scale)
+
+        w2_input_scale = None
+        layer.register_parameter("w2_input_scale", w2_input_scale)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.w13_weight_scale = Parameter(
+            layer.w13_weight_scale.data, requires_grad=False
+        )
+        layer.w2_weight_scale = Parameter(
+            layer.w2_weight_scale.data, requires_grad=False
+        )
+
+        w1_marlin_list = []
+        for e in range(layer.w13_weight.shape[0]):
+            w1_marlin_in = w4a8_2_marlin_weight(layer.w13_weight[e])
+            w1_marlin_list.append(w1_marlin_in)
+        layer.w13_weight = Parameter(torch.stack(w1_marlin_list, dim=0), requires_grad=False)
+        w2_marlin_list = []
+        for e in range(layer.w2_weight.shape[0]):
+            w2_marlin_in = w4a8_2_marlin_weight(layer.w2_weight[e])
+            w2_marlin_list.append(w2_marlin_in)
+        layer.w2_weight = Parameter(torch.stack(w2_marlin_list, dim=0), requires_grad=False)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+        enable_eplb: bool = False,
+        use_nn_moe: Optional[bool] = False,
+        routed_scaling_factor: Optional[float] = None,
+        use_fused_gate: Optional[bool] = False,
+        **_  
+    ) -> torch.Tensor:
+        from vllm.model_executor.layers.fused_moe import fused_experts
+        if enable_eplb:
+            raise NotImplementedError(
+                "EPLB not supported for `SlimQuantW4A8Int8MarlinMoEMethod` yet.")
+        # Expert selection
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias,
+            routed_scaling_factor=routed_scaling_factor,
+            use_fused_gate=use_fused_gate
+        )
+        workspace, global_reduce_buffer = MarlinMoeWorkspace(x.device).get_buffers()
+        return fused_experts_impl_w4a8_marlin(
+            x,
+            layer.w13_weight,
+            layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            workspace=workspace,
+            global_reduce_buffer=global_reduce_buffer,
+            inplace=True,
+            use_int4_w4a8=True,
+            per_channel_quant=True,
+            activation=activation,
+            expert_map=expert_map,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            global_num_experts=global_num_experts,
+            w1_scale=(layer.w13_weight_scale),
+            w2_scale=(layer.w2_weight_scale),
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale,
+            use_nn_moe=use_nn_moe,
+        )

vllm/model_executor/layers/quantization/utils/w4a8_utils.py

+
+import torch
+import numpy as np
+
+def unpack_int8_to_int4(tensor_int8: torch.Tensor) -> torch.Tensor:
+    """
+    将[N, K//2]大小的torch.int8 Tensor，转换为[N, K]大小的torch.int32 Tensor。
+    每个int8包含两个int4，分别提取到int32的低4位，其余位为0。
+
+    Args:
+        tensor_int8 (torch.Tensor): 输入张量，形状为[N, K//2]，类型为torch.int8。
+
+    Returns:
+        torch.Tensor: 输出张量，形状为[N, K]，类型为torch.int32。
+    """
+    if tensor_int8.dtype != torch.int8:
+        raise ValueError("Input tensor must be of type torch.int8")
+
+    N, K_half = tensor_int8.shape
+    tensor_uint8 = tensor_int8.to(torch.uint8)
+    high4 = tensor_uint8 & 0x0F
+    low4 = (tensor_uint8 >> 4) & 0x0F
+    unpacked = torch.empty((N, K_half * 2), dtype=torch.int32, device=tensor_int8.device)
+    unpacked[:, 0::2] = low4.to(torch.int32)
+    unpacked[:, 1::2] = high4.to(torch.int32)
+
+    return unpacked
+
+def get_weight_perms(interleave: bool=True):
+    perm = []
+    for i in range(64):
+
+        for col in range(4):
+            cur_col = (i % 16) * 4 + col
+            for row in range(8):
+                cur_row = (i // 16) * 8 + row
+                cur_idx =  cur_row * 64 + cur_col
+                perm.append(cur_idx)
+
+    perm = np.array(perm)
+    if interleave:
+        interleave = np.array([4, 0, 5, 1, 6, 2, 7, 3])
+        perm = perm.reshape((-1, 8))[:, interleave].ravel()
+
+    perm = torch.from_numpy(perm)
+
+    return perm
+
+def marlin_weights(q_w,weight_perm,k_tile=32,n_tile=64,pack_factor=8):
+    size_k, size_n = q_w.shape
+    q_w = q_w.reshape((size_k // k_tile, k_tile, size_n // n_tile, n_tile))
+    q_w = q_w.permute((0, 2, 1, 3))
+    q_w = q_w.reshape((size_k // k_tile, size_n * k_tile))
+    q_w = q_w.reshape((-1, weight_perm.numel()))[:, weight_perm].reshape(q_w.shape)
+
+    orig_device = q_w.device
+
+    q_w = q_w.cpu().numpy().astype(np.uint32)
+    q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32)
+    for i in range(pack_factor):
+        q_packed |= q_w[:, i::pack_factor] << 4 * i
+    q_packed = torch.from_numpy(q_packed.astype(np.int32)).to(orig_device)
+
+    return q_packed
+
+def w4a8_2_marlin_weight(w4a8_w):
+    full_w4a8_w = unpack_int8_to_int4(w4a8_w)
+    full_w4a8_w = full_w4a8_w.T
+    weight_perm = get_weight_perms()
+    marlin_q_w = marlin_weights(full_w4a8_w, weight_perm, k_tile=32, n_tile=64, pack_factor=8)
+    return marlin_q_w
+

vllm/model_executor/models/deepseek_v2.py

@@ -67,7 +67,7 @@ from .utils import (PPMissingLayer, is_pp_missing_parameter,
 from vllm import _custom_ops as ops
 from vllm.utils import W8a8GetCacheJSON
 
-os.environ['DPSK_FP16_QUICK'] = os.environ.get('DPSK_FP16_QUICK', '1')
+os.environ['DPSK_FP16_QUICK'] = os.environ.get('DPSK_FP16_QUICK', '0')
 class DeepseekV2MLP(nn.Module):
 
     def __init__(
@@ -622,9 +622,13 @@ class DeepseekV2DecoderLayer(nn.Module):
         residual: Optional[torch.Tensor],
     ) -> torch.Tensor:
         # Self Attention
+        
+        # Fix residual FP16 overflow
+        residual_fix_overflow = False
         if residual is None:
             residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
+            residual_fix_overflow = True
         else:
             hidden_states, residual = self.input_layernorm(
                 hidden_states, residual)
@@ -640,7 +644,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             # We scale both hidden_states and residual before
             # rmsnorm, and rmsnorm result would not affect by scale.
             hidden_states *= 1. / self.routed_scaling_factor
-            if self.layer_idx == 0:
+            if self.layer_idx == 0 or residual_fix_overflow:
                 # The residual is shared by all layers, we only scale it on
                 # first layer.
                 residual *= 1. / self.routed_scaling_factor
@@ -778,14 +782,17 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
         self.num_expert_groups = config.n_group
 
         self.moe_layers: list[FusedMoE] = []
+        example_moe = None
         for layer in self.model.layers:
+            if isinstance(layer, PPMissingLayer):
+                continue
+
             assert isinstance(layer, DeepseekV2DecoderLayer)
             if isinstance(layer.mlp, DeepseekV2MoE):
+                example_moe = layer.mlp
                 self.moe_layers.append(layer.mlp.experts)
 
         # Pick last one layer since the first ones may be dense layers.
-        example_moe = typing.cast(
-            DeepseekV2MoE, self.model.layers[config.num_hidden_layers - 1].mlp)
         self.num_logical_experts = example_moe.n_logical_experts
         self.num_physical_experts = example_moe.n_physical_experts
         self.num_local_physical_experts = example_moe.n_local_physical_experts

vllm/platforms/rocm.py

@@ -180,7 +180,7 @@ class RocmPlatform(Platform):
 
     supported_quantization: list[str] = [
         "awq", "gptq", "fp8", "compressed-tensors", "fbgemm_fp8", "gguf",
-        "quark", "ptpc_fp8", "moe_wna16", "blockwise_int8","slimquant_w4a8","awq_marlin"
+        "quark", "ptpc_fp8", "moe_wna16", "blockwise_int8","slimquant_w4a8","awq_marlin","slimquant_w4a8_marlin"
     ]
 
     @classmethod
@@ -282,6 +282,10 @@ class RocmPlatform(Platform):
                 logger.info_once("Using Triton backend on V1 engine.")
                 return TRITON_ATTN_VLLM_V1
             
+        if selected_backend == _Backend.DUAL_CHUNK_FLASH_ATTN:
+            logger.info("Using DualChunkFlashAttention backend.")
+            return ("vllm.attention.backends.dual_chunk_flash_attn."
+                    "DualChunkFlashAttentionBackend")
         if selected_backend == _Backend.ROCM_FLASH:
             if not cls.has_device_capability(90):
                 # not Instinct series GPUs.

vllm/zero_overhead/v1/core.py

@@ -177,11 +177,11 @@ def zero_overhead_update_from_output(scheduler:Scheduler,
     # loop can be a performance bottleneck. We should do our best to avoid
     # expensive operations inside the loop.
     for request in scheduler.running:
+        req_id = request.request_id
         if request.is_finished():
             if req_id in requsets_valid_token_len:
                 requsets_valid_token_len.pop(req_id)
             continue
-        req_id = request.request_id
         num_tokens_scheduled = num_scheduled_tokens.get(req_id, 0)
         if num_tokens_scheduled == 0:
             # The request was not scheduled in this step.