Remove hardcoded device="cuda" to support more devices (#2503)

Co-authored-by: Jiang Li <jiang1.li@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Kunshang Ji
2024-02-01 15:46:39 -08:00
committed by GitHub
co-authored by Jiang Li
parent c410f5d020
commit 96b6f475dd
32 changed files with 343 additions and 292 deletions
@@ -96,7 +96,6 @@ class AWQLinearMethod(LinearMethodBase):
torch.empty(
input_size_per_partition,
output_size_per_partition // self.quant_config.pack_factor,
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -112,7 +111,6 @@ class AWQLinearMethod(LinearMethodBase):
torch.empty(
input_size_per_partition // self.quant_config.group_size,
output_size_per_partition // self.quant_config.pack_factor,
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -128,7 +126,6 @@ class AWQLinearMethod(LinearMethodBase):
torch.empty(
input_size_per_partition // self.quant_config.group_size,
output_size_per_partition,
device="cuda",
dtype=params_dtype,
),
requires_grad=False,
@@ -127,7 +127,6 @@ class GPTQLinearMethod(LinearMethodBase):
torch.empty(
input_size_per_partition // self.quant_config.pack_factor,
output_size_per_partition,
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -145,7 +144,6 @@ class GPTQLinearMethod(LinearMethodBase):
i // self.quant_config.group_size
for i in range(input_size_per_partition)
],
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -156,7 +154,6 @@ class GPTQLinearMethod(LinearMethodBase):
torch.empty(
scale_and_zero_size,
output_size_per_partition // self.quant_config.pack_factor,
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -172,7 +169,6 @@ class GPTQLinearMethod(LinearMethodBase):
torch.empty(
scale_and_zero_size,
output_size_per_partition,
device="cuda",
dtype=params_dtype,
),
requires_grad=False,
@@ -80,7 +80,6 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
torch.empty(
input_size_per_partition // self.quant_config.pack_factor,
output_size_per_partition,
device="cuda",
dtype=torch.int32,
),
requires_grad=False,
@@ -96,7 +95,6 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
torch.empty(
output_size,
self.quant_config.weight_bits**2,
device="cuda",
dtype=params_dtype,
),
requires_grad=False,
@@ -118,12 +116,12 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
out_shape = x.shape[:-1] + (qweight.shape[-1], )
reshaped_x = x.reshape(-1, x.shape[-1])
if is_hip():
out_f = torch.zeros(out_shape, device="cuda", dtype=torch.float)
out_f = torch.zeros(out_shape, dtype=torch.float)
ops.squeezellm_gemm(reshaped_x, qweight, out_f, lookup_table)
out = out_f.to(dtype=torch.float16)
else:
# NOTE: The output tensor should be zero-initialized.
out = torch.zeros(out_shape, device="cuda", dtype=torch.float16)
out = torch.zeros(out_shape, dtype=torch.float16)
ops.squeezellm_gemm(reshaped_x, qweight, out, lookup_table)
if bias is not None: