[Bugfix] Enable padded FP4 quantization (#25947)

roikoren755 · web-flow · commit 4069db3f2e3f · 2025-10-09T10:59:41.000-07:00
Signed-off-by: Roi Koren &lt;roik@nvidia.com&gt;
diff --git a/vllm/_custom_ops.py b/vllm/_custom_ops.py
@@ -1384,7 +1384,7 @@ def scaled_fp4_quant(
     rounded_m = round_up(m, 128)
     scale_n = n // block_size
     rounded_n = round_up(scale_n, 4)
-    output_scale = torch.empty(
+    output_scale = torch.zeros(
         (rounded_m, rounded_n // 4), device=device, dtype=torch.int32
     )
 
diff --git a/vllm/utils/flashinfer.py b/vllm/utils/flashinfer.py
@@ -386,8 +386,6 @@ def flashinfer_scaled_fp4_mm(
     assert block_scale_a.ndim == 2 and block_scale_b.ndim == 2
     assert a.stride(-1) == 1 and b.stride(-1) == 1
     assert a.shape[1] == b.shape[1]
-    assert block_scale_a.shape[1] == a.shape[1] // 8
-    assert block_scale_b.shape[1] == b.shape[1] // 8
 
     if backend == "cutlass":
         block_scale_a = block_scale_a.view(torch.uint8)

Original file line number	Diff line number	Diff line change
`@@ -1384,7 +1384,7 @@ def scaled_fp4_quant(`
`1384`	`1384`	`rounded_m = round_up(m, 128)`
`1385`	`1385`	`scale_n = n // block_size`
`1386`	`1386`	`rounded_n = round_up(scale_n, 4)`
`1387`		`- output_scale = torch.empty(`
	`1387`	`+ output_scale = torch.zeros(`
`1388`	`1388`	`(rounded_m, rounded_n // 4), device=device, dtype=torch.int32`
`1389`	`1389`	`)`
`1390`	`1390`