Remove VLLM_TEST_FORCE_FP8_MARLIN to replace with linear_backend/moe_backend (#52182)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Misha Goin
2026-08-18 12:58:04 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 88b2bff2c6
commit bca7bea240
15 changed files with 46 additions and 66 deletions
+8 -3
View File
@@ -2339,6 +2339,10 @@ class TestFP8Layer(torch.nn.Module):
force_kernel: type[_KernelT] | None = None,
):
super().__init__()
self.input_size_per_partition = weight_shape[1]
self.output_size_per_partition = weight_shape[0]
self.logical_widths = [self.output_size_per_partition]
self.orig_dtype = input_dtype
act_scale_desc = activation_quant_key.scale
weight_scale_desc = weight_quant_key.scale
is_block_wise = act_scale_desc.group_shape.is_per_group()
@@ -2346,11 +2350,12 @@ class TestFP8Layer(torch.nn.Module):
block_size = weight_scale_desc.group_shape.col
weight_scale_shape = weight_shape[0] // block_size
self.weight_scale_inv = torch.rand(
(weight_scale_shape, weight_scale_shape), dtype=torch.float32
(weight_scale_shape, weight_scale_shape),
dtype=torch.float32,
device=device,
)
self.weight = torch.rand(weight_shape).to(dtype=FP8_DTYPE)
self.weight = torch.rand(weight_shape, device=device).to(dtype=FP8_DTYPE)
self.input_scale = None
self.weight_scale = None
self.weight_block_size = [block_size, block_size]
if transpose_weights:
self.weight = self.weight.t()