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

vllm/model_executor/layers/quantization/awq_triton.py

@@ -0,0 +1,320 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+AWQ_TRITON_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
+
+
+@triton.jit
+def awq_dequantize_kernel(
+        qweight_ptr,  # quantized matrix
+        scales_ptr,  # scales, per group
+        zeros_ptr,  # zeros, per group
+        group_size,  # Should always be one of the supported group sizes
+        result_ptr,  # Output matrix
+        num_cols,  # input num cols in qweight
+        num_rows,  # input num rows in qweight
+        BLOCK_SIZE_X: tl.constexpr,
+        BLOCK_SIZE_Y: tl.constexpr):
+    # Setup the pids.
+    pid_x = tl.program_id(axis=0)
+    pid_y = tl.program_id(axis=1)
+
+    # Compute offsets and masks for qweight_ptr.
+    offsets_y = pid_y * BLOCK_SIZE_Y + tl.arange(0, BLOCK_SIZE_Y)
+    offsets_x = pid_x * BLOCK_SIZE_X + tl.arange(0, BLOCK_SIZE_X)
+    offsets = num_cols * offsets_y[:, None] + offsets_x[None, :]
+
+    masks_y = offsets_y < num_rows
+    masks_x = offsets_x < num_cols
+
+    masks = masks_y[:, None] & masks_x[None, :]
+
+    # Compute offsets and masks for result output ptr.
+    result_offsets_y = pid_y * BLOCK_SIZE_Y + tl.arange(0, BLOCK_SIZE_Y)
+    result_offsets_x = pid_x * BLOCK_SIZE_X * 8 + tl.arange(
+        0, BLOCK_SIZE_X * 8)
+    result_offsets = (8 * num_cols * result_offsets_y[:, None] +
+                      result_offsets_x[None, :])
+
+    result_masks_y = result_offsets_y < num_rows
+    result_masks_x = result_offsets_x < num_cols * 8
+    result_masks = result_masks_y[:, None] & result_masks_x[None, :]
+
+    # Load the weights.
+    iweights = tl.load(qweight_ptr + offsets, masks, 0.0)
+    iweights = tl.interleave(iweights, iweights)
+    iweights = tl.interleave(iweights, iweights)
+    iweights = tl.interleave(iweights, iweights)
+
+    # Create reverse AWQ order as tensor: [0, 4, 1, 5, 2, 6, 3, 7]
+    # that will map given indices to the correct order.
+    reverse_awq_order_tensor = ((tl.arange(0, 2) * 4)[None, :] +
+                                tl.arange(0, 4)[:, None]).reshape(8)
+
+    # Use this to compute a set of shifts that can be used to unpack and
+    # reorder the values in iweights and zeros.
+    shifts = reverse_awq_order_tensor * 4
+    shifts = tl.broadcast_to(shifts[None, :], (BLOCK_SIZE_Y * BLOCK_SIZE_X, 8))
+    shifts = tl.reshape(shifts, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Unpack and reorder: shift out the correct 4-bit value and mask.
+    iweights = (iweights >> shifts) & 0xF
+
+    # Compute zero offsets and masks.
+    zero_offsets_y = pid_y * BLOCK_SIZE_Y // group_size + tl.arange(0, 1)
+    zero_offsets_x = pid_x * BLOCK_SIZE_X + tl.arange(0, BLOCK_SIZE_X)
+    zero_offsets = num_cols * zero_offsets_y[:, None] + zero_offsets_x[None, :]
+
+    zero_masks_y = zero_offsets_y < num_rows // group_size
+    zero_masks_x = zero_offsets_x < num_cols
+    zero_masks = zero_masks_y[:, None] & zero_masks_x[None, :]
+
+    # Load the zeros.
+    zeros = tl.load(zeros_ptr + zero_offsets, zero_masks, 0.0)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.broadcast_to(zeros, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Unpack and reorder: shift out the correct 4-bit value and mask.
+    zeros = (zeros >> shifts) & 0xF
+
+    # Compute scale offsets and masks.
+    scale_offsets_y = pid_y * BLOCK_SIZE_Y // group_size + tl.arange(0, 1)
+    scale_offsets_x = (pid_x * BLOCK_SIZE_X * 8 +
+                       tl.arange(0, BLOCK_SIZE_X * 8))
+    scale_offsets = (num_cols * 8 * scale_offsets_y[:, None] +
+                     scale_offsets_x[None, :])
+    scale_masks_y = scale_offsets_y < num_rows // group_size
+    scale_masks_x = scale_offsets_x < num_cols * 8
+    scale_masks = scale_masks_y[:, None] & scale_masks_x[None, :]
+
+    # Load the scales.
+    scales = tl.load(scales_ptr + scale_offsets, scale_masks, 0.0)
+    scales = tl.broadcast_to(scales, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Dequantize.
+    iweights = (iweights - zeros) * scales
+    iweights = iweights.to(result_ptr.type.element_ty)
+
+    # Finally, store.
+    tl.store(result_ptr + result_offsets, iweights, result_masks)
+
+
+@triton.jit
+def awq_gemm_kernel(a_ptr, b_ptr, c_ptr, zeros_ptr, scales_ptr, M, N, K,
+                    group_size, BLOCK_SIZE_M: tl.constexpr,
+                    BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
+                    SPLIT_K: tl.constexpr):
+    pid = tl.program_id(axis=0)
+    pid_z = tl.program_id(1)
+
+    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
+    # num_pid_n = (N + BLOCK_SIZE_N - 1) // BLOCK_SIZE_N
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+
+    pid_m = pid // num_pid_n
+    pid_n = pid % num_pid_n
+
+    accumulator_dtype = c_ptr.type.element_ty
+
+    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
+    # accumulator = tl.arange(0, BLOCK_SIZE_N)
+    # accumulator = tl.broadcast_to(accumulator[None, :],
+    # (BLOCK_SIZE_M, BLOCK_SIZE_N))
+    # accumulator = accumulator & 0x0
+    # accumulator = accumulator.to(accumulator_dtype)
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N),
+                           dtype=accumulator_dtype)
+
+    # Create reverse AWQ order as tensor: [0, 4, 1, 5, 2, 6, 3, 7]
+    # that will map given indices to the correct order.
+    reverse_awq_order_tensor = ((tl.arange(0, 2) * 4)[None, :] +
+                                tl.arange(0, 4)[:, None]).reshape(8)
+
+    # Create the necessary shifts to use to unpack.
+    shifts = reverse_awq_order_tensor * 4
+    shifts = tl.broadcast_to(shifts[None, :],
+                             (BLOCK_SIZE_K * (BLOCK_SIZE_N // 8), 8))
+    shifts = tl.reshape(shifts, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+    # Offsets and masks.
+    offsets_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    masks_am = offsets_am < M
+
+    offsets_bn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
+    masks_bn = offsets_bn < N // 8
+
+    offsets_zn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
+    masks_zn = offsets_zn < N // 8
+
+    offsets_sn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    masks_sn = offsets_sn < N
+
+    offsets_k = pid_z * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
+    offsets_a = K * offsets_am[:, None] + offsets_k[None, :]
+    offsets_b = (N // 8) * offsets_k[:, None] + offsets_bn[None, :]
+
+    a_ptrs = a_ptr + offsets_a
+    b_ptrs = b_ptr + offsets_b
+
+    # NOTE: Use this in TRITON_INTERPRET=1 mode instead of tl.cdiv
+    # block_offset = BLOCK_SIZE_K * SPLIT_K
+    # for k in range(0, (K + block_offset - 1) // (block_offset)):
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)):
+        masks_k = offsets_k < K
+        masks_a = masks_am[:, None] & masks_k[None, :]
+        a = tl.load(a_ptrs, mask=masks_a, other=0.0)
+
+        masks_b = masks_k[:, None] & masks_bn[None, :]
+        b = tl.load(b_ptrs, mask=masks_b, other=0.0)
+        b = tl.interleave(b, b)
+        b = tl.interleave(b, b)
+        b = tl.interleave(b, b)
+
+        # Dequantize b.
+        offsets_szk = (
+            (BLOCK_SIZE_K * SPLIT_K * k + pid_z * BLOCK_SIZE_K) // group_size +
+            tl.arange(0, 1))
+        offsets_z = (N // 8) * offsets_szk[:, None] + offsets_zn[None, :]
+        masks_zk = offsets_szk < K // group_size
+        masks_z = masks_zk[:, None] & masks_zn[None, :]
+        zeros_ptrs = zeros_ptr + offsets_z
+        zeros = tl.load(zeros_ptrs, mask=masks_z, other=0.0)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.broadcast_to(zeros, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+        offsets_s = N * offsets_szk[:, None] + offsets_sn[None, :]
+        masks_sk = offsets_szk < K // group_size
+        masks_s = masks_sk[:, None] & masks_sn[None, :]
+        scales_ptrs = scales_ptr + offsets_s
+        scales = tl.load(scales_ptrs, mask=masks_s, other=0.0)
+        scales = tl.broadcast_to(scales, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+        b = (b >> shifts) & 0xF
+        zeros = (zeros >> shifts) & 0xF
+        b = (b - zeros) * scales
+        b = b.to(c_ptr.type.element_ty)
+
+        # Accumulate results.
+        accumulator = tl.dot(a, b, accumulator, out_dtype=accumulator_dtype)
+
+        offsets_k += BLOCK_SIZE_K * SPLIT_K
+        a_ptrs += BLOCK_SIZE_K * SPLIT_K
+        b_ptrs += BLOCK_SIZE_K * SPLIT_K * (N // 8)
+
+    c = accumulator.to(c_ptr.type.element_ty)
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + pid_z * N * M + N * offs_cm[:, None] + offs_cn[None, :]
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+# qweights - [K     , M // 8], int32
+# scales   - [K // G, M     ], float16
+# zeros    - [K // G, M // 8], int32
+def awq_dequantize_triton(qweight: torch.Tensor,
+                          scales: torch.Tensor,
+                          zeros: torch.Tensor,
+                          block_size_x: int = 32,
+                          block_size_y: int = 32) -> torch.Tensor:
+    K = qweight.shape[0]
+    M = scales.shape[1]
+    group_size = qweight.shape[0] // scales.shape[0]
+
+    assert K > 0 and M > 0
+    assert scales.shape[0] == K // group_size and scales.shape[1] == M
+    assert zeros.shape[0] == K // group_size and zeros.shape[1] == M // 8
+    assert group_size <= K
+    assert group_size in AWQ_TRITON_SUPPORTED_GROUP_SIZES or group_size == K
+
+    # Result tensor:
+    # number of rows = same as input tensor
+    # number of cols = 8 x input tensor num cols
+    result = torch.empty(qweight.shape[0],
+                         qweight.shape[1] * 8,
+                         device=qweight.device,
+                         dtype=scales.dtype)
+
+    Y = qweight.shape[0]  # num rows
+    X = qweight.shape[1]  # num cols
+
+    grid = lambda META: (
+        triton.cdiv(X, META['BLOCK_SIZE_X']),
+        triton.cdiv(Y, META['BLOCK_SIZE_Y']),
+    )
+    awq_dequantize_kernel[grid](qweight,
+                                scales,
+                                zeros,
+                                group_size,
+                                result,
+                                X,
+                                Y,
+                                BLOCK_SIZE_X=block_size_x,
+                                BLOCK_SIZE_Y=block_size_y)
+
+    return result
+
+
+# input   - [M, K]
+# qweight - [K, N // 8]
+# qzeros  - [K // G, N // 8]
+# scales  - [K // G, N]
+# split_k_iters - parallelism along K-dimension, int, power of 2.
+def awq_gemm_triton(input: torch.Tensor,
+                    qweight: torch.Tensor,
+                    scales: torch.Tensor,
+                    qzeros: torch.Tensor,
+                    split_k_iters: int,
+                    block_size_m: int = 32,
+                    block_size_n: int = 32,
+                    block_size_k: int = 32) -> torch.Tensor:
+    M, K = input.shape
+    N = qweight.shape[1] * 8
+    group_size = qweight.shape[0] // qzeros.shape[0]
+
+    assert N > 0 and K > 0 and M > 0
+    assert qweight.shape[0] == K and qweight.shape[1] == N // 8
+    assert qzeros.shape[0] == K // group_size and qzeros.shape[1] == N // 8
+    assert scales.shape[0] == K // group_size and scales.shape[1] == N
+    assert split_k_iters & (split_k_iters - 1) == 0 and split_k_iters != 0
+    assert split_k_iters <= 32
+    assert group_size <= K
+    assert group_size in AWQ_TRITON_SUPPORTED_GROUP_SIZES or group_size == K
+
+    grid = lambda META: (
+        triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(
+            N, META['BLOCK_SIZE_N']),
+        split_k_iters,
+    )
+
+    result = torch.zeros((split_k_iters, M, N),
+                         dtype=scales.dtype,
+                         device=input.device)
+
+    # A = input, B = qweight, C = result
+    # A = M x K, B = K x N, C = M x N
+    awq_gemm_kernel[grid](input,
+                          qweight,
+                          result,
+                          qzeros,
+                          scales,
+                          M,
+                          N,
+                          K,
+                          group_size,
+                          BLOCK_SIZE_M=block_size_m,
+                          BLOCK_SIZE_N=block_size_n,
+                          BLOCK_SIZE_K=block_size_k,
+                          SPLIT_K=split_k_iters)
+
+    result = result.sum(0)
+
+    return result

vllm/model_executor/layers/quantization/base_config.py

@@ -0,0 +1,151 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import inspect
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+from torch import nn
+
+if TYPE_CHECKING:
+    from vllm.model_executor.layers.quantization import QuantizationMethods
+else:
+    QuantizationMethods = str
+
+
+class QuantizeMethodBase(ABC):
+    """Base class for different quantized methods."""
+
+    @abstractmethod
+    def create_weights(self, layer: torch.nn.Module, *weight_args,
+                       **extra_weight_attrs):
+        """Create weights for a layer.
+
+        The weights will be set as attributes of the layer."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply(self, layer: torch.nn.Module, *args, **kwargs) -> torch.Tensor:
+        """Apply the weights in layer to the input tensor.
+
+        Expects create_weights to have been called before on the layer."""
+        raise NotImplementedError
+
+    # Not required functions
+    def embedding(self, layer: torch.nn.Module, *args,
+                  **kwargs) -> torch.Tensor:
+        """Gather embeddings in the layer based on indices in the input tensor.
+
+        Expects create_weights to have been called before on the layer."""
+        raise NotImplementedError
+
+    def process_weights_after_loading(self, layer: nn.Module) -> None:
+        """Process the weight after loading.
+
+        This can be used for example, to transpose weights for computation.
+        """
+        return
+
+
+def method_has_implemented_embedding(
+        method_class: type[QuantizeMethodBase]) -> bool:
+    """
+    Not all quant methods have embedding implemented, so we need to check that
+    it exists for our given method. We check this by making sure the function
+    has been changed from the base implementation.
+    """
+    base_embedding = inspect.getattr_static(QuantizeMethodBase, "embedding",
+                                            None)
+    class_embedding = inspect.getattr_static(method_class, "embedding", None)
+
+    return (class_embedding is not None
+            and class_embedding is not base_embedding)
+
+
+class QuantizationConfig(ABC):
+    """Base class for quantization configs."""
+
+    def __init__(self):
+        super().__init__()
+        # mapping is updated by models as they initialize
+        self.packed_modules_mapping: dict[str, list[str]] = dict()
+
+    @abstractmethod
+    def get_name(self) -> QuantizationMethods:
+        """Name of the quantization method."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_supported_act_dtypes(self) -> list[torch.dtype]:
+        """List of supported activation dtypes."""
+        raise NotImplementedError
+
+    @classmethod
+    @abstractmethod
+    def get_min_capability(cls) -> int:
+        """Minimum GPU capability to support the quantization method.
+
+        E.g., 70 for Volta, 75 for Turing, 80 for Ampere.
+        This requirement is due to the custom CUDA kernels used by the
+        quantization method.
+        """
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_config_filenames() -> list[str]:
+        """List of filenames to search for in the model directory."""
+        raise NotImplementedError
+
+    @classmethod
+    @abstractmethod
+    def from_config(cls, config: dict[str, Any]) -> "QuantizationConfig":
+        """Create a config class from the model's quantization config."""
+        raise NotImplementedError
+
+    @classmethod
+    def override_quantization_method(
+            cls, hf_quant_cfg, user_quant) -> Optional[QuantizationMethods]:
+        """
+           Detects if this quantization method can support a given checkpoint
+           format by overriding the user specified quantization method -- 
+           this method should only be overwritten by subclasses in exceptional 
+           circumstances
+        """
+        return None
+
+    @staticmethod
+    def get_from_keys(config: dict[str, Any], keys: list[str]) -> Any:
+        """Get a value from the model's quantization config."""
+        for key in keys:
+            if key in config:
+                return config[key]
+        raise ValueError(f"Cannot find any of {keys} in the model's "
+                         "quantization config.")
+
+    @staticmethod
+    def get_from_keys_or(config: dict[str, Any], keys: list[str],
+                         default: Any) -> Any:
+        """Get a optional value from the model's quantization config."""
+        try:
+            return QuantizationConfig.get_from_keys(config, keys)
+        except ValueError:
+            return default
+
+    @abstractmethod
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional[QuantizeMethodBase]:
+        """Get the quantize method to use for the quantized layer.
+        
+        Args:
+            layer: The layer for the quant method.
+            prefix: The full name of the layer in the state dict
+        Returns:
+            The quantize method. None if the given layer doesn't support quant
+            method.
+        """
+        raise NotImplementedError
+
+    def get_cache_scale(self, name: str) -> Optional[str]:
+        return None