remove generalized pack routine, save for future PR, plus add removed…

… init file

tests/kernels/test_marlinv2_gemm.py

@@ -11,7 +11,7 @@
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    pack_weights_into_int32, quantize_weights)
+    pack_rows, quantize_weights, pack_weights_into_int32)
 from vllm.platforms import current_platform
 from vllm.scalar_type import ScalarType, scalar_types
 
@@ -68,7 +68,8 @@ def marlinv2_quantize_and_pack(w: torch.Tensor,
         # to match how the kernel applies zps
         ref_zero_points_after_scales=True)
 
-    w_q = pack_weights_into_int32(w_q, wtype)
+    w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape)
+    w_q = w_q.t().contiguous().t() # convert to col major
     w_q_marlinv2 = ops.marlinv2_prepack_B(w_q, wtype)
 
     return w_ref, w_q_marlinv2, w_s, w_zp

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

@@ -1,5 +1,5 @@
 """This file is used for /tests and /benchmarks"""
-from typing import List
+from typing import List, Union
 
 import numpy
 import torch
@@ -229,34 +229,6 @@ def pack_rows(
     return q_res
 
 
-def pack_weights_into_int32(w_q: torch.Tensor,
-                            wtype: ScalarType,
-                            dim: int = 0):
-    orig_device = w_q.device
-
-    # move dim to pack to the end
-    perm = (*[i for i in range(len(w_q.shape)) if i != dim], dim)
-    inv_perm = tuple(perm.index(i) for i in range(len(perm)))
-    w_q_perm = w_q.permute(perm)
-
-    w_q_perm = w_q_perm.cpu().numpy().astype(numpy.uint32)
-    pack_factor = 32 // wtype.size_bits
-    mask = (1 << wtype.size_bits) - 1
-
-    new_shape_perm = list(w_q_perm.shape)
-    new_shape_perm[-1] //= pack_factor
-    assert new_shape_perm[-1] % pack_factor == 0
-
-    w_q_res = numpy.zeros(new_shape_perm, dtype=numpy.uint32)
-    for i in range(pack_factor):
-        w_q_res |= (w_q_perm[..., i::pack_factor]
-                    & mask) << wtype.size_bits * i
-
-    w_q_res = torch.from_numpy(w_q_res.astype(numpy.int32)).to(orig_device)
-    w_q_res = w_q_res.permute(inv_perm)
-
-    return w_q_res
-
 
 def pack_cols(
     q_w: torch.Tensor,