Remove QServe and FBGEMM FP8 quantization (#31109)

sgl-kernel/python/sgl_kernel/__init__.py

@@ -61,8 +61,6 @@ else:
         gptq_gemm,
         gptq_shuffle,
         int8_scaled_mm,
-        qserve_w4a8_per_chn_gemm,
-        qserve_w4a8_per_group_gemm,
         sgl_per_token_group_quant_8bit,
         sgl_per_token_group_quant_fp8,
         sgl_per_token_group_quant_int8,
@@ -192,8 +190,6 @@ else:
         "moe_sum",
         "moe_sum_reduce",
         "prepare_moe_input",
-        "qserve_w4a8_per_chn_gemm",
-        "qserve_w4a8_per_group_gemm",
         "reconstruct_indices_from_tree_mask",
         "rmsnorm",
         "rotary_embedding",

sgl-kernel/python/sgl_kernel/gemm.py

@@ -137,50 +137,6 @@ def sgl_per_token_quant_fp8(
     torch.ops.sgl_kernel.sgl_per_token_quant_fp8.default(input, output_q, output_s)
 
 
-def qserve_w4a8_per_chn_gemm(
-    in_feats: torch.Tensor,
-    kernel: torch.Tensor,
-    wscales: torch.Tensor,
-    ascales: torch.Tensor,
-    w_szs: torch.Tensor,
-    a_ssums: torch.Tensor,
-    out_feats: Optional[torch.Tensor] = None,
-) -> torch.Tensor:
-    if out_feats is None:
-        # NOTE(HandH1998): qserve_w4a8_per_chn_gemm only supports out dtype=torch.float16 now
-        out_feats = torch.empty(
-            (in_feats.shape[0], kernel.shape[0]),
-            device=in_feats.device,
-            dtype=torch.float16,
-        )
-    torch.ops.sgl_kernel.qserve_w4a8_per_chn_gemm.default(
-        in_feats, kernel, wscales, ascales, w_szs, a_ssums, out_feats
-    )
-    return out_feats
-
-
-def qserve_w4a8_per_group_gemm(
-    in_feats: torch.Tensor,
-    kernel: torch.Tensor,
-    zeros: torch.Tensor,
-    scales_i8: torch.Tensor,
-    wscales: torch.Tensor,
-    ascales: torch.Tensor,
-    out_feats: Optional[torch.Tensor] = None,
-) -> torch.Tensor:
-    if out_feats is None:
-        # NOTE(HandH1998): qserve_w4a8_per_group_gemm only supports out dtype=torch.float16 now
-        out_feats = torch.empty(
-            (in_feats.shape[0], kernel.shape[0]),
-            device=in_feats.device,
-            dtype=torch.float16,
-        )
-    torch.ops.sgl_kernel.qserve_w4a8_per_group_gemm.default(
-        in_feats, kernel, zeros, scales_i8, wscales, ascales, out_feats
-    )
-    return out_feats
-
-
 def shuffle_rows(input_tensor, dst2src_map, output_tensor_shape):
     output_tensor = torch.empty(
         output_tensor_shape,