diff --git a/python/sglang/srt/layers/quantization/__init__.py b/python/sglang/srt/layers/quantization/__init__.py index f34a581d657..3e2078c4a4d 100644 --- a/python/sglang/srt/layers/quantization/__init__.py +++ b/python/sglang/srt/layers/quantization/__init__.py @@ -13,7 +13,6 @@ from vllm.model_executor.layers.quantization.deepspeedfp import DeepSpeedFPConfig from vllm.model_executor.layers.quantization.experts_int8 import ExpertsInt8Config from vllm.model_executor.layers.quantization.fbgemm_fp8 import FBGEMMFp8Config -from vllm.model_executor.layers.quantization.fp8 import Fp8Config, Fp8MoEMethod from vllm.model_executor.layers.quantization.gguf import GGUFConfig from vllm.model_executor.layers.quantization.gptq import GPTQConfig from vllm.model_executor.layers.quantization.gptq_marlin import GPTQMarlinConfig @@ -23,6 +22,7 @@ from vllm.model_executor.layers.quantization.tpu_int8 import Int8TpuConfig from sglang.srt.layers.quantization.base_config import QuantizationConfig +from sglang.srt.layers.quantization.fp8 import Fp8Config, Fp8MoEMethod QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = { "aqlm": AQLMConfig, @@ -100,13 +100,13 @@ def fp8_moe_apply( def fp8_get_quant_method(self, layer, prefix): """Enhanced get_quant_method for FP8 config.""" from vllm.model_executor.layers.linear import LinearBase - from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod from vllm.model_executor.layers.quantization.utils.quant_utils import ( is_layer_skipped, ) from sglang.srt.layers.fused_moe_triton.layer import FusedMoE from sglang.srt.layers.linear import UnquantizedLinearMethod + from sglang.srt.layers.quantization.fp8 import Fp8LinearMethod if isinstance(layer, LinearBase): if is_layer_skipped(prefix, self.ignored_layers): diff --git a/python/sglang/srt/layers/quantization/fp8.py b/python/sglang/srt/layers/quantization/fp8.py new file mode 100644 index 00000000000..acdce0b8cbd --- /dev/null +++ b/python/sglang/srt/layers/quantization/fp8.py @@ -0,0 +1,559 @@ +# Adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/model_executor/layers/quantization/fp8.py + +import logging +from typing import Any, Callable, Dict, List, Optional + +import torch +from torch.nn import Module +from torch.nn.parameter import Parameter +from vllm import _custom_ops as ops +from vllm.model_executor.layers.linear import LinearBase +from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod +from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import ( + apply_fp8_marlin_linear, + prepare_fp8_layer_for_marlin, +) +from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped +from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( + all_close_1d, + apply_fp8_linear, + convert_to_channelwise, + cutlass_fp8_supported, + per_tensor_dequantize, + requantize_with_max_scale, +) +from vllm.model_executor.parameter import ModelWeightParameter, PerTensorScaleParameter + +from sglang.srt.layers.fused_moe_triton import ( + FusedMoE, + FusedMoEMethodBase, + FusedMoeWeightScaleSupported, +) +from sglang.srt.layers.linear import LinearMethodBase, UnquantizedLinearMethod +from sglang.srt.layers.quantization.base_config import ( + QuantizationConfig, + QuantizeMethodBase, +) +from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz +from sglang.srt.utils import ( + get_bool_env_var, + is_hip, + print_warning_once, + set_weight_attrs, +) + +ACTIVATION_SCHEMES = ["static", "dynamic"] + +logger = logging.getLogger(__name__) + + +class Fp8Config(QuantizationConfig): + """Config class for FP8.""" + + def __init__( + self, + is_checkpoint_fp8_serialized: bool = False, + activation_scheme: str = "dynamic", + ignored_layers: Optional[List[str]] = None, + ) -> None: + self.is_checkpoint_fp8_serialized = is_checkpoint_fp8_serialized + if is_checkpoint_fp8_serialized: + logger.warning( + "Detected fp8 checkpoint. Please note that the " + "format is experimental and subject to change." + ) + if activation_scheme not in ACTIVATION_SCHEMES: + raise ValueError(f"Unsupported activation scheme {activation_scheme}") + self.activation_scheme = activation_scheme + self.ignored_layers = ignored_layers or [] + + @classmethod + def get_name(cls) -> str: + return "fp8" + + @classmethod + def get_supported_act_dtypes(cls) -> List[torch.dtype]: + return [torch.bfloat16, torch.half] + + @classmethod + def get_min_capability(cls) -> int: + return 80 + + @classmethod + def get_config_filenames(cls) -> List[str]: + return [] + + @classmethod + def from_config(cls, config: Dict[str, Any]) -> "Fp8Config": + quant_method = cls.get_from_keys(config, ["quant_method"]) + is_checkpoint_fp8_serialized = "fp8" in quant_method + activation_scheme = cls.get_from_keys(config, ["activation_scheme"]) + ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None) + return cls( + is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized, + activation_scheme=activation_scheme, + ignored_layers=ignored_layers, + ) + + def get_quant_method( + self, layer: torch.nn.Module, prefix: str + ) -> Optional["QuantizeMethodBase"]: + from vllm.attention.layer import Attention # Avoid circular import + + if isinstance(layer, LinearBase): + if is_layer_skipped(prefix, self.ignored_layers): + return UnquantizedLinearMethod() + return Fp8LinearMethod(self) + elif isinstance(layer, FusedMoE): + return Fp8MoEMethod(self) + elif isinstance(layer, Attention): + return Fp8KVCacheMethod(self) + return None + + def get_scaled_act_names(self) -> List[str]: + return [] + + +class Fp8LinearMethod(LinearMethodBase): + """Linear method for FP8. + Supports loading FP8 checkpoints with static weight scale and + dynamic/static activation scale. + + Also supports loading quantized FP16/BF16 model checkpoints with dynamic + activation scaling. The weight scaling factor will be initialized after + the model weights are loaded. + + Limitations: + 1. Only support per-tensor quantization due to torch._scaled_mm support. + 2. Only support float8_e4m3fn data type due to the limitation of + torch._scaled_mm (https://github.com/pytorch/pytorch/blob/2e48b39603411a41c5025efbe52f89560b827825/aten/src/ATen/native/cuda/Blas.cpp#L854-L856) + + Args: + quant_config: The quantization config. + """ + + def __init__(self, quant_config: Fp8Config): + self.quant_config = quant_config + self.cutlass_fp8_supported = cutlass_fp8_supported() + + # For GPUs that lack FP8 hardware support, we can leverage the Marlin + # kernel for fast weight-only FP8 quantization + self.use_marlin = get_bool_env_var("SGLANG_FORCE_FP8_MARLIN") + # Disable marlin for ROCm + if is_hip(): + self.use_marlin = False + + def create_weights( + self, + layer: torch.nn.Module, + input_size_per_partition: int, + output_partition_sizes: List[int], + input_size: int, + output_size: int, + params_dtype: torch.dtype, + **extra_weight_attrs, + ): + del input_size, output_size + output_size_per_partition = sum(output_partition_sizes) + weight_loader = extra_weight_attrs.get("weight_loader") + + layer.logical_widths = output_partition_sizes + + layer.input_size_per_partition = input_size_per_partition + layer.output_size_per_partition = output_size_per_partition + layer.orig_dtype = params_dtype + + # WEIGHT + weight_dtype = ( + torch.float8_e4m3fn + if self.quant_config.is_checkpoint_fp8_serialized + else params_dtype + ) + + weight = ModelWeightParameter( + data=torch.empty( + output_size_per_partition, input_size_per_partition, dtype=weight_dtype + ), + input_dim=1, + output_dim=0, + weight_loader=weight_loader, + ) + layer.register_parameter("weight", weight) + + # If checkpoint is serialized fp8, load them. + # Otherwise, wait until process_weights_after_loading. + if self.quant_config.is_checkpoint_fp8_serialized: + # WEIGHT SCALE + scale = PerTensorScaleParameter( + data=torch.empty(len(output_partition_sizes), dtype=torch.float32), + weight_loader=weight_loader, + ) + + scale[:] = torch.finfo(torch.float32).min + layer.register_parameter("weight_scale", scale) + + # INPUT ACTIVATION SCALE + if self.quant_config.activation_scheme == "static": + scale = PerTensorScaleParameter( + data=torch.empty(len(output_partition_sizes), dtype=torch.float32), + weight_loader=weight_loader, + ) + + scale[:] = torch.finfo(torch.float32).min + layer.register_parameter("input_scale", scale) + else: + layer.register_parameter("input_scale", None) + + def process_weights_after_loading(self, layer: Module) -> None: + layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False) + # If checkpoint not serialized fp8, quantize the weights. + if not self.quant_config.is_checkpoint_fp8_serialized: + qweight, weight_scale = ops.scaled_fp8_quant(layer.weight, scale=None) + + # If using marlin (w8a16), kernel uses channelwise weights, + # so extend the weight scales to be channelwise. + if self.use_marlin: + assert weight_scale.numel() == 1 + weight_scale = convert_to_channelwise( + weight_scale.expand(len(layer.logical_widths)), layer.logical_widths + ) + + # Update the layer with the new values. + layer.weight = Parameter(qweight.t(), requires_grad=False) + layer.weight_scale = Parameter(weight_scale, requires_grad=False) + layer.input_scale = None + + # If checkpoint is fp8, handle that there are N scales for N + # shards in a fused module + else: + layer.weight_scale = torch.nn.Parameter( + layer.weight_scale.data, requires_grad=False + ) + if self.quant_config.activation_scheme == "static": + layer.input_scale = torch.nn.Parameter( + layer.input_scale.data, requires_grad=False + ) + # If using marlin (w8a16), kernel uses channelwise weights, + # so extend the weight scales to be channelwise. + if self.use_marlin: + weight = layer.weight + weight_scale = convert_to_channelwise( + layer.weight_scale, layer.logical_widths + ) + + # If using w8a8, torch._scaled_mm needs per tensor, so + # requantize the logical shards as a single weight. + else: + # Dequant -> Quant with max scale so we can run per tensor. + weight = layer.weight + weight_scale = layer.weight_scale + + # If ROCm, normalize the weights and scales to e4m3fnuz + if is_hip(): + weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz( + weight=weight, + weight_scale=weight_scale, + input_scale=layer.input_scale, + ) + if input_scale is not None: + layer.input_scale = Parameter(input_scale, requires_grad=False) + + weight_scale, weight = requantize_with_max_scale( + weight=weight, + weight_scale=weight_scale, + logical_widths=layer.logical_widths, + ) + + # Update layer with new values. + layer.weight = Parameter(weight.t(), requires_grad=False) + layer.weight_scale = Parameter(weight_scale, requires_grad=False) + if self.quant_config.activation_scheme == "static": + layer.input_scale = Parameter( + layer.input_scale.max(), requires_grad=False + ) + + if self.use_marlin: + prepare_fp8_layer_for_marlin(layer) + # Activations not quantized for marlin. + del layer.input_scale + + def apply( + self, + layer: torch.nn.Module, + x: torch.Tensor, + bias: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + + if self.use_marlin: + return apply_fp8_marlin_linear( + input=x, + weight=layer.weight, + weight_scale=layer.weight_scale, + workspace=layer.workspace, + size_n=layer.output_size_per_partition, + size_k=layer.input_size_per_partition, + bias=bias, + ) + + return apply_fp8_linear( + input=x, + weight=layer.weight, + weight_scale=layer.weight_scale, + input_scale=layer.input_scale, + bias=bias, + cutlass_fp8_supported=self.cutlass_fp8_supported, + use_per_token_if_dynamic=False, + ) + + +class Fp8MoEMethod(FusedMoEMethodBase): + """MoE method for FP8. + Supports loading FP8 checkpoints with static weight scale and + dynamic/static activation scale. + + Also supports loading quantized FP16/BF16 model checkpoints with dynamic + activation scaling. The weight scaling factor will be initialized after + the model weights are loaded. + + Args: + quant_config: The quantization config. + """ + + def __init__(self, quant_config: Fp8Config): + self.quant_config = quant_config + + def create_weights( + self, + layer: Module, + num_experts: int, + hidden_size: int, + intermediate_size: int, + params_dtype: torch.dtype, + **extra_weight_attrs, + ): + + if self.quant_config.is_checkpoint_fp8_serialized: + params_dtype = torch.float8_e4m3fn + + # WEIGHTS + w13_weight = torch.nn.Parameter( + torch.empty( + num_experts, 2 * intermediate_size, hidden_size, dtype=params_dtype + ), + 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, dtype=params_dtype + ), + requires_grad=False, + ) + layer.register_parameter("w2_weight", w2_weight) + set_weight_attrs(w2_weight, extra_weight_attrs) + + # WEIGHT_SCALES + # Allocate 2 scales for w1 and w3 respectively. + # They will be combined to a single scale after weight loading. + w13_weight_scale = torch.nn.Parameter( + torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False + ) + layer.register_parameter("w13_weight_scale", w13_weight_scale) + + w2_weight_scale = torch.nn.Parameter( + torch.ones(num_experts, dtype=torch.float32), requires_grad=False + ) + layer.register_parameter("w2_weight_scale", w2_weight_scale) + # Add the quantization method used (per tensor/grouped/channel) + # to ensure the weight scales are loaded in properly + extra_weight_attrs.update( + {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value} + ) + # If loading fp8 checkpoint, pass the weight loaders. + # If loading an fp16 checkpoint, do not (we will quantize in + # process_weights_after_loading() + if self.quant_config.is_checkpoint_fp8_serialized: + set_weight_attrs(w13_weight_scale, extra_weight_attrs) + set_weight_attrs(w2_weight_scale, extra_weight_attrs) + + # INPUT_SCALES + if self.quant_config.activation_scheme == "static": + if not self.quant_config.is_checkpoint_fp8_serialized: + raise ValueError( + "Found static activation scheme for checkpoint that " + "was not serialized fp8." + ) + + w13_input_scale = torch.nn.Parameter( + torch.ones(num_experts, dtype=torch.float32), requires_grad=False + ) + layer.register_parameter("w13_input_scale", w13_input_scale) + set_weight_attrs(w13_input_scale, extra_weight_attrs) + + w2_input_scale = torch.nn.Parameter( + torch.ones(num_experts, dtype=torch.float32), requires_grad=False + ) + layer.register_parameter("w2_input_scale", w2_input_scale) + set_weight_attrs(w2_input_scale, extra_weight_attrs) + + else: + layer.w13_input_scale = None + layer.w2_input_scale = None + + def process_weights_after_loading(self, layer: Module) -> None: + + # If checkpoint is fp16, quantize in place. + if not self.quant_config.is_checkpoint_fp8_serialized: + # If ROCm, use float8_e4m3fnuz instead (MI300x HW) + fp8_dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn + w13_weight = torch.empty_like(layer.w13_weight.data, dtype=fp8_dtype) + w2_weight = torch.empty_like(layer.w2_weight.data, dtype=fp8_dtype) + + # Re-initialize w13_scale because we directly quantize + # merged w13 weights and generate a single scaling factor. + layer.w13_weight_scale = torch.nn.Parameter( + torch.ones( + layer.num_experts, dtype=torch.float32, device=w13_weight.device + ), + requires_grad=False, + ) + for expert in range(layer.num_experts): + w13_weight[expert, :, :], layer.w13_weight_scale[expert] = ( + ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :]) + ) + w2_weight[expert, :, :], layer.w2_weight_scale[expert] = ( + ops.scaled_fp8_quant(layer.w2_weight.data[expert, :, :]) + ) + layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False) + layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False) + return + + # If checkpoint is fp8, we need to handle that the + # MoE kernels require single activation scale and single weight + # scale for w13 per expert. + else: + # Fp8 moe kernels require a single activation scale. + # We take the max of all the scales in case they differ. + if self.quant_config.activation_scheme == "static": + if layer.w13_input_scale is None or layer.w2_input_scale is None: + raise ValueError( + "QuantConfig has static quantization, but found " + "activation scales are None." + ) + if not all_close_1d(layer.w13_input_scale) or not all_close_1d( + layer.w2_input_scale + ): + print_warning_once( + "Found input_scales that are not equal for " + "fp8 MoE layer. Using the maximum across experts " + "for each layer. " + ) + layer.w13_input_scale = torch.nn.Parameter( + layer.w13_input_scale.max(), requires_grad=False + ) + layer.w2_input_scale = torch.nn.Parameter( + layer.w2_input_scale.max(), requires_grad=False + ) + # If ROCm, normalize the weights and scales to e4m3fnuz + if is_hip(): + # Normalize the weights and scales + w13_weight, w13_weight_scale, w13_input_scale = ( + normalize_e4m3fn_to_e4m3fnuz( + layer.w13_weight, layer.w13_weight_scale, layer.w13_input_scale + ) + ) + w2_weight, w2_weight_scale, w2_input_scale = ( + normalize_e4m3fn_to_e4m3fnuz( + layer.w2_weight, layer.w2_weight_scale, layer.w2_input_scale + ) + ) + # Reset the parameter + layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False) + layer.w13_weight_scale = torch.nn.Parameter( + w13_weight_scale, requires_grad=False + ) + if w13_input_scale is not None: + layer.w13_input_scale = torch.nn.Parameter( + w13_input_scale, requires_grad=False + ) + layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False) + layer.w2_weight_scale = torch.nn.Parameter( + w2_weight_scale, requires_grad=False + ) + if w2_input_scale is not None: + layer.w2_input_scale = torch.nn.Parameter( + w2_input_scale, requires_grad=False + ) + + # Fp8 moe kernel needs single weight scale for w13 per expert. + # We take the max then dequant and requant each expert. + assert layer.w13_weight_scale is not None + shard_size = layer.intermediate_size_per_partition + max_w13_scales = layer.w13_weight_scale.max(dim=1).values + for expert_id in range(layer.num_experts): + start = 0 + for shard_id in range(2): + dq_weight = per_tensor_dequantize( + layer.w13_weight[expert_id][start : start + shard_size, :], + layer.w13_weight_scale[expert_id][shard_id], + ) + layer.w13_weight[expert_id][start : start + shard_size, :], _ = ( + ops.scaled_fp8_quant(dq_weight, max_w13_scales[expert_id]) + ) + start += shard_size + + layer.w13_weight_scale = torch.nn.Parameter( + max_w13_scales, requires_grad=False + ) + return + + def apply( + self, + layer: torch.nn.Module, + x: torch.Tensor, + router_logits: torch.Tensor, + top_k: int, + renormalize: bool, + use_grouped_topk: bool, + topk_group: Optional[int] = None, + num_expert_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + ) -> torch.Tensor: + + from vllm.model_executor.layers.fused_moe import fused_experts + + 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, + ) + + return fused_experts( + x, + layer.w13_weight, + layer.w2_weight, + topk_weights=topk_weights, + topk_ids=topk_ids, + inplace=True, + use_fp8_w8a8=True, + w1_scale=layer.w13_weight_scale, + w2_scale=layer.w2_weight_scale, + a1_scale=layer.w13_input_scale, + a2_scale=layer.w2_input_scale, + ) + + +class Fp8KVCacheMethod(BaseKVCacheMethod): + """ + Supports loading kv-cache scaling factors from FP8 checkpoints. + """ + + def __init__(self, quant_config: Fp8Config): + super().__init__(quant_config)