Update flashinfer to 0.6.8 (#39959)

Signed-off-by: bai <v@gor.io>
This commit is contained in:
bai
2026-04-20 10:37:23 -07:00
committed by GitHub
parent b9cf629bd0
commit 191e3fdaa1
9 changed files with 31 additions and 61 deletions
@@ -62,7 +62,7 @@ def test_supports_batch_invariant_disables():
@patch("vllm.envs.VLLM_BATCH_INVARIANT", False)
@patch(
"vllm.utils.flashinfer.current_platform.is_device_capability",
"vllm.utils.flashinfer.current_platform.is_device_capability_family",
return_value=True,
)
@patch("vllm.utils.flashinfer.has_nvidia_artifactory", return_value=True)
+17 -15
View File
@@ -548,7 +548,9 @@ def test_trtllm_gen_mxfp4_fused_moe(
hidden_states, hidden_states_scale = mxfp8_quantize(
hidden_states, is_sf_swizzled_layout=False
)
hidden_states_scale = hidden_states_scale.view(torch.float8_e4m3fn).reshape(-1)
hidden_states_scale = hidden_states_scale.view(torch.float8_e4m3fn).reshape(
*hidden_states.shape[:-1], -1
)
else:
hidden_states_scale = None
@@ -595,20 +597,20 @@ def test_trtllm_gen_mxfp4_fused_moe(
if beta is not None:
beta = torch.full((num_experts,), beta, device=hidden_states.device)
tg_result = tg_mxfp4_moe(
router_logits,
topk,
num_experts,
intermediate_size,
hidden_size,
hidden_states,
hidden_states_scale,
w13,
w13_scale,
bias13,
w2,
w2_scale,
bias2,
act_type,
router_logits=router_logits,
topk=topk,
num_experts=num_experts,
intermediate_size=intermediate_size,
hidden_size=hidden_size,
hidden_states=hidden_states,
hidden_states_scale=hidden_states_scale,
w13_weight=w13,
w13_weight_scale=w13_scale,
w13_bias=bias13,
w2_weight=w2,
w2_weight_scale=w2_scale,
w2_bias=bias2,
act_type=act_type,
alpha=alpha,
beta=beta,
limit=limit,