fix test_deepep_moe.py

Signed-off-by: Bill Nell <bnell@redhat.com>

Author: bnellnm

tests/kernels/moe/test_deepep_moe.py

@@ -104,10 +104,6 @@ def make(config: TestConfig, low_latency_mode: bool) -> "TestTensors":
         rank_tokens = torch.randn(
             (config.m, config.k), device="cuda", dtype=token_dtype) / 10
         rank_token_scales = None
-        if config.dtype == torch.float8_e4m3fn:
-            # low_latency_mode kernels dont support per-token quant.
-            _, rank_token_scales = ops.scaled_fp8_quant(
-                rank_tokens, use_per_token_if_dynamic=not low_latency_mode)
 
         topk = torch.randint(low=0,
                              high=config.num_experts,

vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py

@@ -157,7 +157,7 @@ def prepare(
                 a1,
                 a1_scale,
                 quant_dtype=quant_config.quant_dtype,
-                per_act_token_quant=False,
+                per_act_token_quant=True,
                 block_shape=quant_config.block_shape,
             )
             (expert_x, expert_x_scale, expert_num_tokens, expert_topk_ids,