sglang 0.4.9.post2__py3-none-any.whl → 0.4.9.post4__py3-none-any.whl

sglang/srt/layers/quantization/utils.py

@@ -1,20 +1,21 @@
 # Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/utils/quant_utils.py
 
+from __future__ import annotations
+
+import re
+from copy import deepcopy
 from types import MappingProxyType
-from typing import List, Mapping, Tuple, Union
+from typing import TYPE_CHECKING, Dict, List, Mapping, Optional, Tuple, Union
 
+import numpy
 import torch
 
 from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
-from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_npu
-
-_is_cuda = is_cuda()
-_is_npu = is_npu()
-_is_cpu_amx_available = cpu_has_amx_support()
-_is_cpu = is_cpu()
+from sglang.srt.layers.quantization.scalar_type import ScalarType, scalar_types
+from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_hip, is_npu
 
-if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available)):
-    from vllm._custom_ops import scaled_fp8_quant
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.base_config import QuantizationConfig
 
 
 def is_layer_skipped(
@@ -143,3 +144,333 @@ def replace_parameter(
         if not isinstance(new, torch.nn.Parameter):
             new = torch.nn.Parameter(new, requires_grad=False)
         mod.register_parameter(name, torch.nn.Parameter(new, requires_grad=False))
+
+
+# Match dynamic rules with module name (prefix) and override quantize
+# config if module (prefix) matches a rule
+def override_config(config: QuantizationConfig, prefix: str):
+    weight_bits = get_dynamic_override(config, prefix, "bits", config.weight_bits)
+    if isinstance(weight_bits, int):
+        config.weight_bits = weight_bits
+    group_size = get_dynamic_override(config, prefix, "group_size", config.group_size)
+    if isinstance(group_size, int):
+        config.group_size = group_size
+    desc_act = get_dynamic_override(config, prefix, "desc_act", config.desc_act)
+    if isinstance(desc_act, bool):
+        config.desc_act = desc_act
+
+    config.pack_factor = 32 // config.weight_bits  # packed into int32
+    if config.get_name() == "gptq_marlin":
+        is_sym = get_dynamic_override(config, prefix, "sym", config.is_sym)
+        if isinstance(is_sym, bool):
+            config.is_sym = is_sym
+
+        if (config.weight_bits, config.is_sym) not in config.TYPE_MAP:
+            raise ValueError(
+                "Unsupported quantization config: "
+                f"bits={config.weight_bits}, sym={config.is_sym}"
+            )
+
+        config.quant_type = config.TYPE_MAP[(config.weight_bits, config.is_sym)]
+    elif config.get_name() == "gptq":
+        if config.weight_bits not in [2, 3, 4, 8]:
+            raise ValueError(
+                "Currently, only 2/3/4/8-bit weight quantization is "
+                f"supported for GPTQ, but got {config.weight_bits} bits."
+            )
+
+
+def get_dynamic_override(
+    config: QuantizationConfig,
+    layer_name: str,
+    key: Optional[str] = None,
+    default_value: Union[int, bool, None] = None,
+) -> Union[Dict, int, bool, None]:
+    for pattern, pattern_dict in config.dynamic.items():
+        # Negative match: matched modules are excluded from quantized init
+        if pattern.startswith("-:"):
+            if re.match(pattern.removeprefix("-:"), layer_name):
+                return False
+        # Positive match: matched modules have quant properties overrides
+        # base quant config
+        elif re.match(pattern.removeprefix("+:"), layer_name):
+            if key is None:
+                return pattern_dict
+            else:
+                return pattern_dict.get(key, default_value)
+    return default_value
+
+
+def get_linear_quant_method(
+    config: QuantizationConfig,
+    layer: torch.nn.Module,
+    prefix: str,
+    linear_method_cls: type,
+):
+    from sglang.srt.layers.linear import LinearBase
+    from sglang.srt.layers.quantization.unquant import (
+        UnquantizedEmbeddingMethod,
+        UnquantizedLinearMethod,
+    )
+    from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+    cloned_config = deepcopy(config)
+    parallel_lm_head_quantized = (
+        isinstance(layer, ParallelLMHead) and cloned_config.lm_head_quantized
+    )
+
+    if isinstance(layer, LinearBase) or parallel_lm_head_quantized:
+        # False = skip module, None = no override, else = Positive match
+        if get_dynamic_override(cloned_config, layer_name=prefix) is False:
+            if parallel_lm_head_quantized:
+                return UnquantizedEmbeddingMethod()
+            return UnquantizedLinearMethod()
+
+        if prefix:
+            # Dynamic per module/layer rules may override base config
+            override_config(cloned_config, prefix=prefix)
+
+        return linear_method_cls(cloned_config)
+    return None
+
+
+def get_pack_factor(num_bits):
+    assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}"
+    return 32 // num_bits
+
+
+def permute_rows(
+    q_w: torch.Tensor,
+    w_ref: torch.Tensor,
+    group_size: int,
+    test_perm: Optional[torch.Tensor] = None,
+):
+    assert q_w.shape == w_ref.shape
+
+    orig_device = q_w.device
+    k_size, _ = q_w.shape
+
+    g_idx = torch.zeros((k_size,), dtype=torch.int32)
+    for i in range(k_size):
+        g_idx[i] = i // group_size
+
+    # Simulate act_order by doing a random permutation on K
+    rand_perm = test_perm if test_perm is not None else torch.randperm(k_size)
+
+    g_idx = g_idx[rand_perm].contiguous()
+    q_w = q_w[rand_perm, :].contiguous()
+    w_ref = w_ref[rand_perm, :].contiguous()
+
+    return (
+        w_ref.to(device=orig_device),
+        q_w.to(device=orig_device),
+        g_idx.to(device=orig_device),
+        rand_perm.to(device=orig_device),
+    )
+
+
+def pack_cols(
+    q_w: torch.Tensor,
+    num_bits: int,
+    size_k: int,
+    size_n: int,
+):
+    assert q_w.shape == (size_k, size_n)
+
+    pack_factor = get_pack_factor(num_bits)
+    assert size_n % pack_factor == 0
+
+    orig_device = q_w.device
+
+    q_w = q_w.cpu().numpy().astype(numpy.uint32)
+
+    q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32)
+
+    for i in range(pack_factor):
+        q_res |= q_w[:, i::pack_factor] << num_bits * i
+
+    q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device)
+    q_res = q_res.contiguous()
+
+    return q_res
+
+
+def unpack_cols(
+    packed_q_w: torch.Tensor,
+    num_bits: int,
+    size_k: int,
+    size_n: int,
+):
+    pack_factor = get_pack_factor(num_bits)
+    assert size_n % pack_factor == 0
+    assert packed_q_w.shape == (
+        size_k,
+        size_n // pack_factor,
+    ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format(
+        packed_q_w.shape, size_k, size_n, pack_factor
+    )
+
+    orig_device = packed_q_w.device
+
+    packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32)
+    q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32)
+
+    mask = (1 << num_bits) - 1
+    for i in range(pack_factor):
+        vals = packed_q_w_cpu & mask
+        packed_q_w_cpu >>= num_bits
+        q_res[:, i::pack_factor] = vals
+
+    q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device)
+    q_res = q_res.contiguous()
+
+    return q_res
+
+
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/utils/quant_utils.py
+def quantize_weights(
+    w: torch.Tensor,
+    quant_type: ScalarType,
+    group_size: Optional[int],
+    zero_points: bool = False,
+    ref_zero_points_after_scales: bool = False,
+):
+    assert (
+        quant_type.is_integer()
+    ), "Floating point quantization may work but has not been tested"
+    assert not zero_points or group_size is not None, (
+        "to have group zero points, group_size must be provided "
+        "(-1 group_size is channelwise)"
+    )
+
+    orig_device = w.device
+    orig_type = w.dtype
+    size_k, size_n = w.shape
+
+    assert w.is_floating_point(), "w must be float"
+
+    if group_size == -1:
+        group_size = size_k
+
+    # Reshape to [groupsize, -1]
+    if group_size is not None and group_size < size_k:
+        w = w.reshape((-1, group_size, size_n))
+        w = w.permute(1, 0, 2)
+        w = w.reshape((group_size, -1))
+
+    # Compute scale for each group
+    max_val = torch.max(w, 0, keepdim=True).values
+    min_val = torch.min(w, 0, keepdim=True).values
+
+    max_q_val = quant_type.max()
+    min_q_val = quant_type.min()
+
+    w_s = torch.Tensor([1.0]).to(w.device)  # unscaled case
+    maybe_w_zp = None
+    if group_size is not None:
+        if zero_points:
+            assert not quant_type.is_signed() and quant_type.max() > 0
+            w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max()
+            maybe_w_zp = (
+                torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int()
+            )
+        else:
+            # If the bias is such that there are no possible negative/positive
+            #  values, set the max value to inf to avoid divide by 0
+            w_s = torch.max(
+                abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)),
+                abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)),
+            )
+
+    # Quantize
+    w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0)
+    w_q = torch.clamp(w_q, min_q_val, max_q_val)
+
+    # Compute ref (dequantized)
+    # For some kernels (namely Machete) the zero-points are applied after the
+    # scales are applied, for this case computing the reference in similar way
+    # allows us to use tighter error tolerances in our unit tests.
+    if ref_zero_points_after_scales and maybe_w_zp is not None:
+        w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s
+    else:
+        w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s
+
+    if quant_type.has_bias():
+        w_q += quant_type.bias
+
+    # Restore original shapes
+    if group_size is not None and group_size < size_k:
+
+        def reshape_w(w):
+            w = w.reshape((group_size, -1, size_n))
+            w = w.permute(1, 0, 2)
+            w = w.reshape((size_k, size_n)).contiguous()
+            return w
+
+        w_q = reshape_w(w_q)
+        w_ref = reshape_w(w_ref)
+        w_s = w_s.reshape((-1, size_n)).contiguous()
+
+    if maybe_w_zp is not None:
+        maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous()
+        maybe_w_zp = maybe_w_zp.to(device=orig_device)
+
+    return (
+        w_ref.to(device=orig_device),
+        w_q.to(device=orig_device),
+        w_s if group_size is not None else None,
+        maybe_w_zp,
+    )
+
+
+SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128]
+SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
+
+
+def gptq_quantize_weights(
+    w: torch.Tensor,
+    quant_type: ScalarType,
+    group_size: int,
+    act_order: bool,
+    test_perm: Optional[torch.Tensor] = None,
+):
+    size_k, _ = w.shape
+
+    assert w.is_floating_point(), "w must be float"
+    assert (
+        quant_type in SUPPORTED_GPTQ_QUANT_TYPES
+    ), f"Unsupported gptq type = {quant_type}"
+    assert group_size in SUPPORTED_GROUP_SIZES + [
+        size_k
+    ], f"Unsupported groupsize = {group_size}"
+
+    w_ref, w_q, w_s, _ = quantize_weights(w, quant_type, group_size)
+
+    # Apply act_order
+    g_idx = torch.empty(0, dtype=torch.int, device=w.device)
+    rand_perm = torch.empty(0, dtype=torch.int, device=w.device)
+    if act_order:
+        assert (
+            group_size < size_k
+        ), "For act_order, groupsize = {} must be less than size_k = {}".format(
+            group_size, size_k
+        )
+
+        w_ref, w_q, g_idx, rand_perm = permute_rows(w_q, w_ref, group_size, test_perm)
+
+    return w_ref, w_q, w_s, g_idx, rand_perm
+
+
+def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor):
+    orig_device = q_w.device
+
+    sort_indices = torch.argsort(g_idx).to(dtype=torch.int32)  # Sort based on g_idx
+
+    g_idx = g_idx[sort_indices].contiguous()
+    q_w = q_w[sort_indices, :].contiguous()
+
+    return (
+        q_w.to(device=orig_device),
+        g_idx.to(device=orig_device),
+        sort_indices.to(device=orig_device),
+    )

sglang/srt/layers/quantization/w4afp8.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import logging
 from typing import Any, Dict, List, Optional
 
@@ -5,12 +7,13 @@ import torch
 from torch.nn import Module
 from torch.nn.parameter import Parameter
 
-from sglang.srt.layers.linear import LinearBase, UnquantizedLinearMethod
 from sglang.srt.layers.quantization.base_config import (
+    FusedMoEMethodBase,
     QuantizationConfig,
     QuantizeMethodBase,
 )
 from sglang.srt.layers.quantization.fp8 import Fp8LinearMethod
+from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
 from sglang.srt.layers.quantization.utils import is_layer_skipped
 from sglang.srt.utils import set_weight_attrs
 
@@ -62,7 +65,7 @@ class W4AFp8Config(QuantizationConfig):
         return []
 
     @classmethod
-    def from_config(cls, config: Dict[str, Any]) -> "W4AFp8Config":
+    def from_config(cls, config: Dict[str, Any]) -> W4AFp8Config:
         quant_method = cls.get_from_keys(config, ["quant_method"])
         is_checkpoint_fp8_serialized = "fp8" in quant_method
         is_checkpoint_w4afp8_serialized = "w4afp8" in quant_method
@@ -79,7 +82,8 @@ class W4AFp8Config(QuantizationConfig):
 
     def get_quant_method(
         self, layer: torch.nn.Module, prefix: str
-    ) -> Optional["QuantizeMethodBase"]:
+    ) -> Optional[QuantizeMethodBase]:
+        from sglang.srt.layers.linear import LinearBase
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 
         if isinstance(layer, LinearBase):
@@ -94,7 +98,7 @@ class W4AFp8Config(QuantizationConfig):
         return []
 
 
-class W4AFp8MoEMethod:
+class W4AFp8MoEMethod(FusedMoEMethodBase):
 
     def __init__(self, quant_config: W4AFp8Config):
         self.quant_config = quant_config

sglang/srt/layers/quantization/w8a8_fp8.py

@@ -1,11 +1,14 @@
-from typing import Any, Callable, Dict, List, Optional
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import torch
 from torch.nn.parameter import Parameter
 
-from sglang.srt.layers.linear import LinearMethodBase
 from sglang.srt.layers.parameter import ChannelQuantScaleParameter, ModelWeightParameter
 from sglang.srt.layers.quantization.base_config import (
+    FusedMoEMethodBase,
+    LinearMethodBase,
     QuantizationConfig,
     QuantizeMethodBase,
 )
@@ -22,6 +25,9 @@ from sglang.srt.layers.quantization.fp8_utils import (
 )
 from sglang.srt.utils import set_weight_attrs
 
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.topk import TopKOutput
+
 _is_fp8_fnuz = is_fp8_fnuz()
 
 
@@ -64,7 +70,7 @@ class W8A8Fp8Config(QuantizationConfig):
         return []
 
     @classmethod
-    def from_config(cls, config: Dict[str, Any]) -> "W8A8Fp8Config":
+    def from_config(cls, config: Dict[str, Any]) -> W8A8Fp8Config:
         quant_method = cls.get_from_keys(config, ["quant_method"])
         is_checkpoint_fp8_serialized = (
             "compressed-tensors" in quant_method or "w8a8_fp8" in quant_method
@@ -75,7 +81,7 @@ class W8A8Fp8Config(QuantizationConfig):
         self,
         layer: torch.nn.Module,
         prefix: str,
-    ) -> Optional["QuantizeMethodBase"]:
+    ) -> Optional[QuantizeMethodBase]:
         from sglang.srt.layers.linear import LinearBase
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 
@@ -183,7 +189,7 @@ class W8A8Fp8LinearMethod(LinearMethodBase):
         )
 
 
-class W8A8FP8MoEMethod:
+class W8A8FP8MoEMethod(FusedMoEMethodBase):
     """MoE method for FP8.
     Supports loading FP8 checkpoints with static weight scale and
     dynamic/static activation scale.
@@ -194,25 +200,7 @@ class W8A8FP8MoEMethod:
         quant_config: The quantization config.
     """
 
-    def __new__(cls, *args, **kwargs):
-        from sglang.srt.layers.moe.fused_moe_triton import FusedMoEMethodBase
-
-        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):
+    def __init__(self, quant_config: W8A8Fp8Config):
         self.quant_config = quant_config
 
     def create_weights(
@@ -281,45 +269,23 @@ class W8A8FP8MoEMethod:
         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,
-        num_fused_shared_experts: int = 0,
-        custom_routing_function: Optional[Callable] = None,
-        correction_bias: Optional[torch.Tensor] = None,
+        topk_output: TopKOutput,
+        *,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
         routed_scaling_factor: Optional[float] = None,
     ) -> torch.Tensor:
         from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
-        from sglang.srt.layers.moe.topk import select_experts
-
-        # Expert selection
-        topk_weights, topk_ids = 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,
-            num_fused_shared_experts=num_fused_shared_experts,
-            custom_routing_function=custom_routing_function,
-            correction_bias=correction_bias,
-            routed_scaling_factor=routed_scaling_factor,
-        )
 
         return fused_experts(
             x,
             layer.w13_weight,
             layer.w2_weight,
-            topk_weights=topk_weights,
-            topk_ids=topk_ids,
+            topk_output=topk_output,
             inplace=inplace,
+            apply_router_weight_on_input=apply_router_weight_on_input,
             activation=activation,
             use_fp8_w8a8=True,
             per_channel_quant=True,