[Kernel] Optimize FP8 support for MoE kernel / Mixtral via static scales

csrc/ops.h

@@ -146,7 +146,12 @@ void gptq_shuffle(
   torch::Tensor q_perm,
   int bit);
 
-void scaled_fp8_quant(
+void static_scaled_fp8_quant(
+  torch::Tensor& out,
+  torch::Tensor& input,
+  torch::Tensor& scale);
+
+void dynamic_scaled_fp8_quant(
   torch::Tensor& out,
   torch::Tensor& input,
   torch::Tensor& scale);

csrc/pybind.cpp

@@ -73,7 +73,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   ops.def("gptq_gemm", &gptq_gemm, "Quantized GEMM for GPTQ");
   ops.def("gptq_shuffle", &gptq_shuffle, "Post processing for GPTQ");
   ops.def("squeezellm_gemm", &squeezellm_gemm, "Quantized GEMM for SqueezeLLM");
-  ops.def("scaled_fp8_quant", &scaled_fp8_quant, "Compute FP8 quantized tensor and scaling factor");
+  ops.def("static_scaled_fp8_quant", &static_scaled_fp8_quant, "Compute FP8 quantized tensor for given scaling factor");
+  ops.def("dynamic_scaled_fp8_quant", &dynamic_scaled_fp8_quant, "Compute FP8 quantized tensor and scaling factor");
   ops.def(
     "moe_align_block_size",
     &moe_align_block_size,

csrc/quantization/fp8/fp8_cuda_kernels.cu

@@ -74,7 +74,30 @@ __global__ void scaled_fp8_quant_kernel(
 
 } // namespace vllm
 
-void scaled_fp8_quant(
+void static_scaled_fp8_quant(
+  torch::Tensor& out,      // [..., d]
+  torch::Tensor& input,    // [..., d]
+  torch::Tensor& scale)    // [1]
+{
+  int64_t num_tokens = input.numel() / input.size(-1);
+  int64_t num_elems = input.numel();
+  dim3 grid(num_tokens);
+  dim3 block(1024);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(
+    input.scalar_type(),
+    "scaled_fp8_quant_kernel",
+    [&] {
+      vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        out.data_ptr<c10::Float8_e4m3fn>(),
+        input.data_ptr<scalar_t>(),
+        scale.data_ptr<float>(),
+        num_elems);
+      });
+}
+
+void dynamic_scaled_fp8_quant(
   torch::Tensor& out,      // [..., d]
   torch::Tensor& input,    // [..., d]
   torch::Tensor& scale)    // [1]

vllm/_custom_ops.py

@@ -154,10 +154,16 @@ def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
 
 
 # fp8
-def scaled_fp8_quant(input: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    scale = torch.zeros(1, device=input.device, dtype=torch.float32)
+def scaled_fp8_quant(
+    input: torch.Tensor,
+    scale: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
     output = torch.empty_like(input, dtype=torch.float8_e4m3fn)
-    vllm_ops.scaled_fp8_quant(output, input, scale)
+    if scale is None:
+        scale = torch.zeros(1, device=input.device, dtype=torch.float32)
+        vllm_ops.dynamic_scaled_fp8_quant(output, input, scale)
+    else:
+        vllm_ops.static_scaled_fp8_quant(output, input, scale)
     return output, scale
 
 

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -220,8 +220,9 @@ def moe_align_block_size(
 
 
 def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
-                            B_scale: torch.Tensor, topk_weights: torch.Tensor,
-                            topk_ids: torch.Tensor,
+                            A_scale: Optional[torch.Tensor],
+                            B_scale: Optional[torch.Tensor],
+                            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
                             sorted_token_ids: torch.Tensor,
                             expert_ids: torch.Tensor,
                             num_tokens_post_padded: torch.Tensor,
@@ -232,10 +233,10 @@ def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
     assert sorted_token_ids.stride(0) == 1
 
     if not use_fp8:
-        A_scale = None
+        assert A_scale is None
         assert B_scale is None
     else:
-        A, A_scale = ops.scaled_fp8_quant(A)
+        A, A_scale = ops.scaled_fp8_quant(A, A_scale)
         assert B_scale is not None
 
     grid = lambda META: (triton.cdiv(sorted_token_ids.shape[0], META[
@@ -318,6 +319,8 @@ def fused_moe(
     use_fp8: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     """
     This function computes a Mixture of Experts (MoE) layer using two sets of
@@ -434,6 +437,7 @@ def fused_moe(
     invoke_fused_moe_kernel(hidden_states,
                             w1,
                             intermediate_cache1,
+                            a1_scale,
                             w1_scale,
                             topk_weights,
                             topk_ids,
@@ -451,6 +455,7 @@ def fused_moe(
     invoke_fused_moe_kernel(intermediate_cache2,
                             w2,
                             intermediate_cache3,
+                            a2_scale,
                             w2_scale,
                             topk_weights,
                             topk_ids,

vllm/model_executor/layers/quantization/base_config.py

@@ -29,6 +29,12 @@ def get_min_capability(self) -> int:
         """
         raise NotImplementedError
 
+    # The following is not an abstract method and returns True by default.
+    @classmethod
+    def require_config_file(cls) -> bool:
+        """Whether this quantization config needs a configuration filen."""
+        return True
+
     @staticmethod
     @abstractmethod
     def get_config_filenames() -> List[str]:

vllm/model_executor/layers/quantization/fp8.py

@@ -13,6 +13,12 @@
 class FP8Config(QuantizationConfig):
     """Config class for FP8."""
 
+    def __init__(
+        self,
+        act_scaling: str = "dynamic",
+    ) -> None:
+        self.act_scaling = act_scaling
+
     @classmethod
     def get_name(cls) -> str:
         return "fp8"
@@ -28,13 +34,18 @@ def get_min_capability(cls) -> int:
         # be included: https://github.com/pytorch/pytorch/pull/118881
         return 90
 
+    @classmethod
+    def require_config_file(cls) -> bool:
+        return False
+
     @classmethod
     def get_config_filenames(cls) -> List[str]:
-        return []
+        return ["quantize_config.json"]
 
     @classmethod
     def from_config(cls, config: Dict[str, Any]) -> "FP8Config":
-        return cls()
+        act_scaling = cls.get_from_keys(config, ["act_scaling"])
+        return cls(act_scaling)
 
     def get_linear_method(self) -> "Fp8LinearMethod":
         return Fp8LinearMethod(self)

vllm/model_executor/model_loader/weight_utils.py

@@ -135,18 +135,19 @@ def get_quant_config(model_config: ModelConfig,
     else:
         hf_folder = model_name_or_path
 
-    possible_config_filenames = quant_cls.get_config_filenames()
-
-    # If the quantization config is not found, use the default config.
-    if not possible_config_filenames:
-        return quant_cls()
-
     config_files = glob.glob(os.path.join(hf_folder, "*.json"))
+    possible_config_filenames = quant_cls.get_config_filenames()
 
     quant_config_files = [
         f for f in config_files if any(
             f.endswith(x) for x in possible_config_filenames)
     ]
+
+    # If the quantization config is optional and not provided, use the
+    # default config.
+    if not quant_cls.require_config_file() and not quant_config_files:
+        return quant_cls()
+
     if len(quant_config_files) == 0:
         raise ValueError(
             f"Cannot find the config file for {model_config.quantization}")

vllm/model_executor/models/mixtral.py

@@ -104,6 +104,13 @@ def __init__(
                         device="cuda",
                         dtype=self.params_dtype))
 
+        set_weight_attrs(self.ws, {
+            "weight_loader": self.weight_loader,
+        })
+        set_weight_attrs(self.w2s, {
+            "weight_loader": self.weight_loader,
+        })
+
         # Scaling factors for FP8 weights
         self.ws_scale = nn.Parameter(
             torch.ones(
@@ -114,12 +121,23 @@ def __init__(
                 self.num_total_experts, device="cuda", dtype=torch.float32),
             requires_grad=False) if self.use_fp8 else None
 
-        set_weight_attrs(self.ws, {
-            "weight_loader": self.weight_loader,
-        })
-        set_weight_attrs(self.w2s, {
-            "weight_loader": self.weight_loader,
-        })
+        # Scaling factors for FP8 activations
+        need_act_scales = (self.use_fp8 and
+                           linear_method.quant_config.act_scaling == "static")
+        self.as_scale = nn.Parameter(
+            torch.zeros(1, device="cuda", dtype=torch.float32),
+            requires_grad=False) if need_act_scales else None
+        self.a2s_scale = nn.Parameter(
+            torch.zeros(1, device="cuda", dtype=torch.float32),
+            requires_grad=False) if need_act_scales else None
+
+        if need_act_scales:
+            set_weight_attrs(self.as_scale, {
+                "weight_loader": self.weight_loader,
+            })
+            set_weight_attrs(self.a2s_scale, {
+                "weight_loader": self.weight_loader,
+            })
 
     def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
                       weight_name: str, expert_id: int):
@@ -134,6 +152,8 @@ def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
                        shard_size:2 * shard_size, :] = loaded_weight[shard, :]
         if weight_name.endswith("w2.weight"):
             param_data[expert_id, :, :] = loaded_weight[:, shard]
+        if "activation_scale" in weight_name:
+            param_data[:] = param_data[:].max(loaded_weight)
 
     def process_weights_after_loading(self):
         if self.use_fp8:
@@ -161,7 +181,9 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
                                         inplace=True,
                                         use_fp8=self.use_fp8,
                                         w1_scale=self.ws_scale,
-                                        w2_scale=self.w2s_scale)
+                                        w2_scale=self.w2s_scale,
+                                        a1_scale=self.as_scale,
+                                        a2_scale=self.a2s_scale)
 
         if self.tp_size > 1:
             final_hidden_states = tensor_model_parallel_all_reduce(
@@ -443,11 +465,19 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         ]
 
         expert_params_mapping = [
+            # These are the weights for the experts
             # (param_name, weight_name, expert_id)
             ("ws" if weight_name in ["w1", "w3"] else "w2s",
              f"experts.{expert_id}.{weight_name}.weight", expert_id)
             for expert_id in range(self.config.num_local_experts)
             for weight_name in ["w1", "w2", "w3"]
+        ] + [
+            # These are the activation scales for the experts
+            # (param_name, weight_name, expert_id)
+            ("as_scale" if weight_name in ["w1", "w3"] else "a2s_scale",
+             f"experts.{expert_id}.{weight_name}.activation_scale", expert_id)
+            for expert_id in range(self.config.num_local_experts)
+            for weight_name in ["w1", "w2", "w3"]
         ]
 
         params_dict = dict(self.named_parameters())