vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py

@@ -0,0 +1,55 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from abc import ABC, abstractmethod
+from typing import Optional
+
+import torch
+
+__all__ = ["QuarkScheme"]
+
+
+class QuarkScheme(ABC):
+    """
+    Abstract class used to describe the weight creation and forward pass 
+    of different quantization schemes supported by Quark.
+    """
+
+    @classmethod
+    @abstractmethod
+    def get_min_capability(cls) -> int:
+        """
+        Get minimum device capability.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        Weight creation for the particular scheme. Inputs to this function 
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
+                      bias: Optional[torch.Tensor]):
+        """
+        Run the forward pass for the particular scheme. This is where 
+        scheme-specific dequant/quant steps/kernels should be applied.
+
+        :param layer: torch.nn.Module with the registered weights and 
+            other parameters relevant to the particular scheme. 
+        :param x: input to the layer
+        :param bias: bias parameter
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError

vllm/model_executor/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py

@@ -0,0 +1,126 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Any, Callable, Optional
+
+import torch
+import torch.nn.functional as F
+
+import vllm.envs as envs
+from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
+from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
+    OCP_MX_BLOCK_SIZE, per_token_group_quant_mxfp4)
+from vllm.model_executor.parameter import (GroupQuantScaleParameter,
+                                           PackedvLLMParameter)
+from vllm.platforms import current_platform
+
+__all__ = ["QuarkW4A4MXFP4"]
+
+
+class QuarkW4A4MXFP4(QuarkScheme):
+
+    def __init__(self, weight_quant_spec: dict[str, Any],
+                 input_quant_spec: dict[str, Any]):
+        self.out_dtype = torch.get_default_dtype()
+        self.qscheme = "per_group"
+        self.weight_quant_spec = weight_quant_spec
+        self.input_quant_spec = input_quant_spec
+        self.emulate = not current_platform.supports_mx()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.weight = torch.nn.Parameter(layer.weight.data,
+                                          requires_grad=False)
+        layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data,
+                                                requires_grad=False)
+
+        if self.emulate:
+            try:
+                from quark.torch.export.nn.modules import realquantizer
+                from quark.torch.quantization.config.config import (
+                    QuantizationSpec)
+            except ImportError as err:
+                raise ImportError(
+                    "The package `amd-quark` is required to use AMD Quark "
+                    "MX-FP4 models. Please install it with `pip install "
+                    "amd-quark`.") from err
+
+            weight_quant_spec = QuantizationSpec.from_dict(
+                self.weight_quant_spec)
+
+            weight_quantizer = realquantizer.get_real_quantizer(
+                qspec=weight_quant_spec,
+                quantizer=None,
+                real_quantized=True,
+                reorder=False,
+                float_dtype=self.out_dtype,
+                scale_shape=layer.weight_scale.shape,
+                zero_point_shape=None,
+            )
+            weight_quantizer.scale.data = layer.weight_scale.data
+
+            if not envs.VLLM_QUARK_EMU_MEM_OPT:
+                layer.weight = torch.nn.Parameter(
+                    weight_quantizer(layer.weight.data).to(self.out_dtype),
+                    requires_grad=False,
+                )
+            else:
+                self.weight_quantizer = weight_quantizer
+            layer.weight_scale = None
+
+            # This call is necessary to release the scales memory.
+            torch.cuda.empty_cache()
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: list[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+
+        # WEIGHT
+        weight = PackedvLLMParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition // 2,
+                dtype=torch.uint8,
+            ),
+            input_dim=1,
+            output_dim=0,
+            packed_dim=1,
+            packed_factor=2,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        weight_scale = GroupQuantScaleParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition // OCP_MX_BLOCK_SIZE,
+                dtype=torch.uint8,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight_scale", weight_scale)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+        if self.emulate:
+            if envs.VLLM_QUARK_EMU_MEM_OPT:
+                dq_w = self.weight_quantizer(layer.weight).to(self.out_dtype)
+            else:
+                dq_w = layer.weight
+            qdq_x, _ = per_token_group_quant_mxfp4(x, OCP_MX_BLOCK_SIZE)
+            return F.linear(qdq_x, dq_w, bias)
+        else:
+            raise NotImplementedError()

vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py

@@ -0,0 +1,146 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Callable, Optional
+
+import torch
+from torch.nn import Parameter
+
+from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    Fp8LinearOp, normalize_e4m3fn_to_e4m3fnuz, requantize_with_max_scale)
+from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+from vllm.platforms import current_platform
+
+__all__ = ["QuarkW8A8Fp8"]
+
+
+class QuarkW8A8Fp8(QuarkScheme):
+
+    def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool]):
+        self.qscheme = qscheme
+        self.is_static_input_scheme = is_static_input_scheme
+        self.fp8_linear = Fp8LinearOp(use_per_token_if_dynamic=False)
+        self.out_dtype = torch.get_default_dtype()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # lovelace and up
+        return 89
+
+    def process_weights_after_loading(self, layer) -> None:
+        # If per tensor, when we have a fused module (e.g. QKV) with per
+        # tensor scales (thus N scales being passed to the kernel),
+        # requantize so we can always run per tensor
+        if self.qscheme == "per_tensor":
+            if current_platform.is_rocm():
+                input_scale = getattr(layer, 'input_scale', None)
+                weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=layer.weight,
+                    weight_scale=layer.weight_scale,
+                    input_scale=input_scale)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+            else:
+                max_w_scale = layer.weight_scale
+                weight = layer.weight
+
+            max_w_scale, weight = requantize_with_max_scale(
+                weight=weight,
+                weight_scale=max_w_scale,
+                logical_widths=layer.logical_widths,
+            )
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
+
+        # If channelwise, scales are already lined up, so just transpose.
+        elif self.qscheme == "per_channel":
+            weight = layer.weight
+
+            if current_platform.is_fp8_fnuz():
+                input_scale = getattr(layer, 'input_scale', None)
+                weight, weight_scale, input_scale = \
+                    normalize_e4m3fn_to_e4m3fnuz(
+                        weight=weight,
+                        weight_scale=layer.weight_scale,
+                        input_scale=input_scale)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+            else:
+                weight_scale = layer.weight_scale.data
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+        else:
+            raise ValueError(f"Unknown quantization scheme {self.qscheme}")
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            layer.input_scale = Parameter(layer.input_scale.max(),
+                                          requires_grad=False)
+        else:
+            layer.input_scale = None
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: list[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+
+        # WEIGHT
+        weight = ModelWeightParameter(data=torch.empty(
+            output_size_per_partition,
+            input_size_per_partition,
+            dtype=torch.float8_e4m3fn),
+                                      input_dim=1,
+                                      output_dim=0,
+                                      weight_loader=weight_loader)
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        # TODO: update create_xxx_parameter functions to return
+        # the newly added parameters
+        if self.qscheme == "per_channel":
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes)),
+                                 dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader)
+        else:
+            assert self.qscheme == "per_tensor"
+            weight_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                   weight_loader=weight_loader)
+
+        # min requirement for fp8 kernels
+        weight_scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                  weight_loader=weight_loader)
+            input_scale[:] = torch.finfo(torch.float32).min
+            layer.register_parameter("input_scale", input_scale)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+        return self.fp8_linear.apply(input=x,
+                                     weight=layer.weight,
+                                     weight_scale=layer.weight_scale,
+                                     out_dtype=self.out_dtype,
+                                     input_scale=layer.input_scale,
+                                     bias=bias)

vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py

@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Callable, Optional
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.kernels.scaled_mm import (
+    ScaledMMLinearLayerConfig, choose_scaled_mm_linear_kernel)
+from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
+from vllm.model_executor.parameter import (BasevLLMParameter,
+                                           ChannelQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+
+logger = init_logger(__name__)
+
+
+class QuarkW8A8Int8(QuarkScheme):
+    _kernel_backends_being_used: set[str] = set()
+
+    def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool],
+                 input_symmetric: Optional[bool]):
+        self.qscheme = qscheme
+        self.is_static_input_scheme = is_static_input_scheme
+        self.input_symmetric = input_symmetric
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # turing and up
+        return 75
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: list[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+        layer.logical_widths = output_partition_sizes
+
+        scaled_mm_linear_kernel_config = ScaledMMLinearLayerConfig(
+            is_channelwise=(self.qscheme == "per_channel"),
+            is_static_input_scheme=(self.is_static_input_scheme is True),
+            input_symmetric=(self.input_symmetric is True))
+
+        kernel_type = choose_scaled_mm_linear_kernel(
+            scaled_mm_linear_kernel_config)
+
+        if kernel_type.__name__ not in self._kernel_backends_being_used:
+            logger.info("Using %s for QuarkW8A8Int8", kernel_type.__name__)
+            self._kernel_backends_being_used.add(kernel_type.__name__)
+
+        # WEIGHT
+        weight = ModelWeightParameter(data=torch.empty(
+            sum(output_partition_sizes),
+            input_size_per_partition,
+            dtype=torch.int8),
+                                      input_dim=1,
+                                      output_dim=0,
+                                      weight_loader=weight_loader)
+
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        if self.qscheme == "per_channel":
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes)),
+                                 dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader)
+            ChannelQuantZPParameter = ChannelQuantScaleParameter
+            weight_zero_point = ChannelQuantZPParameter(
+                data=torch.empty((sum(output_partition_sizes)),
+                                 dtype=torch.int8),
+                output_dim=0,
+                weight_loader=weight_loader)
+        else:
+            assert self.qscheme == "per_tensor"
+            weight_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                   weight_loader=weight_loader)
+            PerTensorZPParameter = PerTensorScaleParameter
+            weight_zero_point = PerTensorZPParameter(
+                data=torch.empty(len(output_partition_sizes),
+                                 dtype=torch.int8),
+                weight_loader=weight_loader)
+        layer.register_parameter("weight_scale", weight_scale)
+        layer.register_parameter("weight_zero_point", weight_zero_point)
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = BasevLLMParameter(data=torch.empty(
+                1, dtype=torch.float32),
+                                            weight_loader=weight_loader)
+            layer.register_parameter("input_scale", input_scale)
+
+            input_zero_point = BasevLLMParameter(data=torch.empty(
+                1, dtype=torch.int8),
+                                                 weight_loader=weight_loader)
+            layer.register_parameter("input_zero_point", input_zero_point)
+
+        self.kernel = kernel_type(c=scaled_mm_linear_kernel_config,
+                                  w_q_param_name="weight",
+                                  w_s_param_name="weight_scale",
+                                  i_s_param_name="input_scale",
+                                  i_zp_param_name="input_zero_point",
+                                  azp_adj_param_name="azp_adj")
+
+    # Checkpoints are serialized in quark format, which is
+    # different from the format the kernel may want. Handle repacking here.
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.register_parameter("weight_zero_point", None)
+        delattr(layer, 'weight_zero_point')
+        if self.input_symmetric:
+            layer.register_parameter("input_zero_point", None)
+            delattr(layer, 'input_zero_point')
+
+        self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
+                      bias: Optional[torch.Tensor]) -> torch.Tensor:
+        return self.kernel.apply_weights(layer, x, bias)

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

@@ -0,0 +1,105 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Iterable, Mapping
+from types import MappingProxyType
+from typing import Any, Optional
+
+import regex as re
+
+
+def deep_compare(dict1: Any, dict2: Any) -> bool:
+    if type(dict1) is not type(dict2):
+        return False
+    if isinstance(dict1, dict):
+        if dict1.keys() != dict2.keys():
+            return False
+        return all(deep_compare(dict1[k], dict2[k]) for k in dict1)
+    elif isinstance(dict1, list):
+        return set(dict1) == set(dict2)
+    else:
+        return dict1 == dict2
+
+
+def should_ignore_layer(
+    layer_name: Optional[str],
+    ignore: Iterable[str],
+    fused_mapping: Mapping[str, list[str]] = MappingProxyType({})
+) -> bool:
+    if layer_name is None:
+        return False
+
+    # layer_name = model.layers.0.self_attn.qkv_proj
+    # proj_name = qkv_proj
+    proj_name = layer_name.split(".")[-1]
+
+    # Fused layers like gate_up_proj or qkv_proj will not be fused
+    # in the safetensors checkpoint. So, we convert the name
+    # from the fused version to unfused + check to make sure that
+    # each shard of the fused layer has the same scheme.
+    if proj_name in fused_mapping:
+        shard_proj_names = fused_mapping[proj_name]
+
+        # Convert fused_name --> [shard_names]
+        shard_names = [
+            layer_name.replace(proj_name, shard_proj_name)
+            for shard_proj_name in shard_proj_names
+        ]
+
+        # Layer should be ignored if shards are ignored.
+        should_ignore_layer = None
+        for shard_name in shard_names:
+            should_ignore_shard = check_equal_or_regex_match(
+                layer_name=shard_name, targets=ignore)
+
+            # If shard_idx=0, set layer ignore to match shard.
+            if should_ignore_layer is None:
+                should_ignore_layer = should_ignore_shard
+
+            # If shard_idx=1+ confirm scheme matches prior shards.
+            elif should_ignore_shard != should_ignore_layer:
+                raise ValueError(f"Found a different quantization schemes for "
+                                 f"{shard_proj_names} in {layer_name}. vLLM "
+                                 "requires all to use the same scheme.")
+
+    # Unfused layers like down_proj and o_proj will match
+    # the safetensors checkpoint already.
+    else:
+        should_ignore_layer = check_equal_or_regex_match(layer_name=layer_name,
+                                                         targets=ignore)
+
+    assert should_ignore_layer is not None
+    return should_ignore_layer
+
+
+def check_equal_or_regex_match(layer_name: str,
+                               targets: Iterable[str]) -> bool:
+    """
+    Checks whether a layer_name is exactly equal or a regex match for 
+    if target starts with 're:' to any target in list.
+    """
+    for target in targets:
+        if _is_equal_or_regex_match(layer_name, target):
+            return True
+    return False
+
+
+def _is_equal_or_regex_match(value: str,
+                             target: str,
+                             check_contains: bool = False) -> bool:
+    """
+    Checks whether a value is exactly equal or a regex match for target
+    if target starts with 're:'. If check_contains is set to True,
+    additionally checks if the target string is contained within the value.
+    """
+
+    if target.startswith("re:"):
+        pattern = target[3:]
+        if re.match(pattern, value):
+            return True
+    elif check_contains:
+        if target.lower() in value.lower():
+            return True
+    elif target == value:
+        return True
+    return False

vllm/model_executor/layers/quantization/schema.py

@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+This file contains the Pydantic schemas for various quantization-related
+parameters. When a relevant quantization technique is specified, these
+parameters are loaded in the form of a JSON alongside the model weights
+and augment the model with additional information needed for use of that
+technique. The format of this JSON should be specified by one or more
+schemas contained here.
+
+For example, when the KV cache is quantized to FP8-E4M3 (currently only
+possible on ROCm), the model can be optionally augmented with KV cache
+scaling factors.
+"""
+
+from typing import Optional
+
+from pydantic import BaseModel, ConfigDict, ValidationInfo, model_validator
+
+
+class KVCacheQuantSchema(BaseModel):
+    dtype: str
+    # Each key is a TP rank. Each value is a dictionary mapping a TP rank's
+    # layer indices to their per-tensor KV cache scaling factor.
+    # TODO: Consider pulling this and its validation methods out into its
+    # own schema class (tricky as its members are variable)
+    scaling_factor: dict[int, dict[int, float]]
+
+    @model_validator(mode="after")
+    def check_is_fp8(self) -> "KVCacheQuantSchema":
+        assert self.dtype == "float8_e4m3fn", (
+            "Loaded scaling factors intended for KV cache dtype = "
+            f"{self.dtype} rather than float8_e4m3fn!")
+        return self
+
+    @model_validator(mode="after")
+    def check_tp_ranks(self, info: ValidationInfo) -> "KVCacheQuantSchema":
+        context = info.context
+        if context:
+            tp_size = context["tp_size"]
+            num_hidden_layers = context["num_hidden_layers"]
+            assert len(self.scaling_factor) == tp_size, (
+                f"Loaded dictionary has TP size {len(self.scaling_factor)} "
+                f"but LLM engine is currently running with TP size {tp_size}.")
+            for tp_rank, layer_maps in self.scaling_factor.items():
+                assert len(layer_maps) == num_hidden_layers, (
+                    f"KV cache scales map for TP rank {tp_rank} is malformed. "
+                    f"Expected {num_hidden_layers} layers, got "
+                    f"{len(layer_maps)}.")
+            for i in range(tp_size):
+                assert i in self.scaling_factor, (
+                    f"KV cache scales map for TP rank {i} not found.")
+        return self
+
+    @model_validator(mode="after")
+    def check_current_rank(self, info: ValidationInfo) -> "KVCacheQuantSchema":
+        context = info.context
+        if context:
+            tp_rank = context["tp_rank"]
+            num_hidden_layers = context["num_hidden_layers"]
+            layer_scales_map = self.scaling_factor[tp_rank]
+            for i in range(num_hidden_layers):
+                assert i in layer_scales_map, (
+                    f"Could not find KV cache scales for layer {i} in "
+                    f"TP rank {tp_rank}.")
+        return self
+
+
+class QuantParamSchema(BaseModel):
+    # TODO: Generalize and extend with more fields
+    # (e.g. weights/activations params) once functionality is enabled
+    model_config = ConfigDict(protected_namespaces=())
+    model_type: Optional[str]
+    kv_cache: KVCacheQuantSchema
+
+    @model_validator(mode="after")
+    def check_model_type(self, info: ValidationInfo) -> "QuantParamSchema":
+        context = info.context
+        if context:
+            model_type = context.get("model_type", None)
+            if model_type is not None:
+                assert model_type == self.model_type, (
+                    f"Model type is {model_type} but loaded "
+                    f"scaling factors belonging to different "
+                    f"model type {self.model_type}!")
+        return self