[Kernel Slimming] Remove sgl-kernel AOT marlin kernels (#19241)

sgl-kernel/python/sgl_kernel/__init__.py

@@ -48,7 +48,6 @@ from sgl_kernel.gemm import (
     fp8_blockwise_scaled_mm,
     fp8_scaled_mm,
     gptq_gemm,
-    gptq_marlin_gemm,
     gptq_shuffle,
     int8_scaled_mm,
     qserve_w4a8_per_chn_gemm,
@@ -78,11 +77,6 @@ from sgl_kernel.mamba import (
     causal_conv1d_update_cpu,
     chunk_gated_delta_rule_cpu,
 )
-from sgl_kernel.marlin import (
-    awq_marlin_moe_repack,
-    awq_marlin_repack,
-    gptq_marlin_repack,
-)
 from sgl_kernel.memory import set_kv_buffer_kernel, weak_ref_tensor
 from sgl_kernel.moe import (
     apply_shuffle_mul_sum,

sgl-kernel/python/sgl_kernel/gemm.py

@@ -1,7 +1,6 @@
 from typing import Optional, Tuple
 
 import torch
-from sgl_kernel.scalar_type import ScalarType
 from sgl_kernel.utils import _get_cache_buf
 
 
@@ -506,46 +505,6 @@ def scaled_fp4_experts_quant(
 
 
 # GPTQ kernels
-def gptq_marlin_gemm(
-    a: torch.Tensor,
-    c: Optional[torch.Tensor],
-    b_q_weight: torch.Tensor,
-    b_scales: torch.Tensor,
-    global_scale: Optional[torch.Tensor],
-    b_zeros: Optional[torch.Tensor],
-    g_idx: Optional[torch.Tensor],
-    perm: Optional[torch.Tensor],
-    workspace: torch.Tensor,
-    b_q_type: ScalarType,
-    size_m: int,
-    size_n: int,
-    size_k: int,
-    is_k_full: bool = True,
-    use_atomic_add: bool = False,
-    use_fp32_reduce: bool = False,
-    is_zp_float: bool = False,
-) -> torch.Tensor:
-    return torch.ops.sgl_kernel.gptq_marlin_gemm(
-        a,
-        c,
-        b_q_weight,
-        b_scales,
-        global_scale,
-        b_zeros,
-        g_idx,
-        perm,
-        workspace,
-        b_q_type.id,
-        size_m,
-        size_n,
-        size_k,
-        is_k_full,
-        use_atomic_add,
-        use_fp32_reduce,
-        is_zp_float,
-    )
-
-
 def gptq_gemm(
     a: torch.Tensor,
     b_q_weight: torch.Tensor,

sgl-kernel/python/sgl_kernel/marlin.py

@@ -1,44 +0,0 @@
-import torch
-
-
-def gptq_marlin_repack(
-    b_q_weight,
-    perm,
-    size_k,
-    size_n,
-    num_bits,
-) -> torch.Tensor:
-    return torch.ops.sgl_kernel.gptq_marlin_repack(
-        b_q_weight,
-        perm,
-        size_k,
-        size_n,
-        num_bits,
-    )
-
-
-def awq_marlin_repack(
-    b_q_weight: torch.Tensor, size_k: int, size_n: int, num_bits: int
-) -> torch.Tensor:
-    return torch.ops.sgl_kernel.awq_marlin_repack(b_q_weight, size_k, size_n, num_bits)
-
-
-def awq_marlin_moe_repack(
-    b_q_weight: torch.Tensor,
-    perm: torch.Tensor,
-    size_k: int,
-    size_n: int,
-    num_bits: int,
-) -> torch.Tensor:
-    num_experts = b_q_weight.shape[0]
-    assert size_k % 16 == 0
-    output = torch.empty(
-        (num_experts, size_k // 16, size_n * (num_bits // 2)),
-        device=b_q_weight.device,
-        dtype=b_q_weight.dtype,
-    )
-    for e in range(num_experts):
-        output[e] = torch.ops.sgl_kernel.awq_marlin_repack(
-            b_q_weight[e], size_k, size_n, num_bits
-        )
-    return output