[CI] Add E2E Blackwell Quantized MoE Test (#25723)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-09-26 12:23:00 -07:00
committed by GitHub
parent 0002b7f0d1
commit f708bd4904
4 changed files with 155 additions and 2 deletions
@@ -40,6 +40,8 @@ def flashinfer_fused_moe_blockscale_fp8(
assert global_num_experts % 4 == 0
assert top_k < (topk_group * global_num_experts / num_expert_group)
assert block_shape == [128, 128]
# Routing kernel expects #experts <= #threads 256
assert global_num_experts <= 256
a_q, a_sf = per_token_group_quant_fp8(x, block_shape[1])
# NOTE: scales of hidden states have to be transposed!