[CI/Build] Bump flashinfer to v0.6.11.post2 (#41711)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
This commit is contained in:
Artem Perevedentsev
2026-05-16 14:55:12 -07:00
committed by GitHub
co-authored by Vadim Gimpelson
parent 36e74c9ea4
commit 0867497368
6 changed files with 30 additions and 23 deletions
+6 -2
View File
@@ -453,11 +453,15 @@ def test_flashinfer_cutedsl_moe_masked(
)
# reference
a_fp4, a_scale_interleaved = fp4_quantize(hidden_states, input_global_scale)
# input_global_scale is per-expert ([num_experts]); fp4_quantize and
# dequantize_nvfp4_to_dtype are non-grouped APIs that expect [1] or
# [num_tokens]. Use a single element since all values are uniform here.
a_global = input_global_scale[:1].contiguous()
a_fp4, a_scale_interleaved = fp4_quantize(hidden_states, a_global)
a_in_dtype = dequantize_nvfp4_to_dtype(
a_fp4,
a_scale_interleaved,
input_global_scale,
a_global,
dtype=hidden_states.dtype,
device=hidden_states.device,
block_size=16,