fixes

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

vllm/distributed/device_communicators/all2all.py

@@ -83,6 +83,9 @@ class PPLXAll2AllManager(All2AllManagerBase):
         assert has_pplx, "pplx_kernels not found. Please follow https://github.com/vllm-project/vllm/blob/main/tools/ep_kernels/README.md to install pplx_kernels."  # noqa
         super().__init__(cpu_group)
 
+        # Intranode doesn't work yet.
+        self.internode = True
+
         if self.internode:
             # inter-node communication needs nvshmem,
             # intra-node communication uses p2p mapping directly

vllm/model_executor/layers/fused_moe/layer.py

@@ -269,9 +269,12 @@ class FusedMoEMethodBase(QuantizeMethodBase):
                 hidden_dim_scale_bytes=(0 if moe.in_dtype.itemsize != 1 else (
                     (moe.hidden_dim + moe.block_size - 1) // moe.block_size *
                     torch.float32.itemsize)),
-                group_name=all2all_manager.cpu_group.group_name,
             )
 
+            if not all2all_manager.internode:
+                all_to_all_args["group_name"] = \
+                    all2all_manager.cpu_group.group_name
+
             handle = all2all_manager.get_handle(all_to_all_args)
 
             logger.debug("PplxPrepareAndFinalize")

vllm/model_executor/layers/quantization/fp8.py

@@ -790,7 +790,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 max_num_tokens=MOE_DP_CHUNK_SIZE,
                 world_size=all2all_manager.world_size,
                 dp_size=all2all_manager.tp_group.world_size,
-                qtype=torch.float8_e4m3fn,
+                use_fp8_w8a8=True,
                 block_shape=self.quant_config.weight_block_size,
                 per_act_token=False,  #?
             )