Merge pull request #102 from s572915912/s572915912-patch-1

【犀牛鸟实战issue】training: split_sizes error

hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinol518-bf16-lr1e4-4096.yaml

@@ -58,6 +58,11 @@ model:
     first_stage_config:
       target: hy3dshape.models.autoencoders.ShapeVAE
       from_pretrained: tencent/Hunyuan3D-2.1
+      #Ensure same with the dataset setting
+      params: 
+        pc_size: 81920              
+        pc_sharpedge_size: 0 
+
 
     cond_stage_config:
       target: hy3dshape.models.conditioner.SingleImageEncoder

hy3dshape/hy3dshape/utils/misc.py

@@ -49,10 +49,12 @@ def instantiate_from_config(config, **kwargs):
     cls = get_obj_from_str(config["target"])
 
     if config.get("from_pretrained", None):
+        params_kwargs = config.get("params", {})
         return cls.from_pretrained(
                     config["from_pretrained"], 
                     use_safetensors=config.get('use_safetensors', False),
-                    variant=config.get('variant', 'fp16'))
+                    variant=config.get('variant', 'fp16'),
+                    **params_kwargs) 
 
     params = config.get("params", dict())
     # params.update(kwargs)

hy3dshape/scripts/train_deepspeed.sh

@@ -1,11 +1,22 @@
 # If: ImportError: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.20' not found 
 # Do: ln /usr/local/gcc-8.3/lib64/libstdc++.so.6 -sf /usr/lib64/libstdc++.so.6
 
+# Try to auto-detect a suitable network interface if NCCL_SOCKET_IFNAME is not already set.
+if [ -z "$NCCL_SOCKET_IFNAME" ]; then
+    # Find the first physical-like interface by excluding common virtual/loopback names.
+    DETECTED_IFACE=$(ls /sys/class/net | grep -vE '^(lo|docker|veth|cali|tunl|kube|ib|usb)' | head -n 1)
+    if [ -n "$DETECTED_IFACE" ]; then
+        echo "NCCL_SOCKET_IFNAME is not set. Auto-detected and exporting: $DETECTED_IFACE"
+        export NCCL_SOCKET_IFNAME=$DETECTED_IFACE
+    else
+        echo "Warning: Could not auto-detect a network interface. You may need to set NCCL_SOCKET_IFNAME manually if NCCL fails."
+    fi
+fi
+
 export NCCL_IB_TIMEOUT=24
 export NCCL_NVLS_ENABLE=0
-NET_TYPE="high"
+NET_TYPE="high" 
 if [[ "${NET_TYPE}" = "low" ]]; then
-    export NCCL_SOCKET_IFNAME=eth1
     export NCCL_IB_GID_INDEX=3
     export NCCL_IB_HCA=mlx5_2:1,mlx5_2:1
     export NCCL_IB_SL=3
@@ -21,9 +32,6 @@ else
     export NCCL_IB_DISABLE=0
     export NCCL_LL_THRESHOLD=16384
     export NCCL_IB_CUDA_SUPPORT=1
-    export NCCL_SOCKET_IFNAME=bond1
-    export UCX_NET_DEVICES=bond1
-    export NCCL_IB_HCA=mlx5_bond_1,mlx5_bond_5,mlx5_bond_3,mlx5_bond_7,mlx5_bond_4,mlx5_bond_8,mlx5_bond_2,mlx5_bond_6
     export NCCL_COLLNET_ENABLE=0
     export SHARP_COLL_ENABLE_SAT=0
     export NCCL_NET_GDR_LEVEL=2
@@ -40,12 +48,13 @@ master_ip=$4
 config=$5
 output_dir=$6
 
-
-echo node_num $node_num
-echo node_rank $node_rank
-echo master_ip $master_ip
-echo config $config
-echo output_dir $output_dir
+echo "--- Script Arguments ---"
+echo "node_num: $node_num"
+echo "node_rank: $node_rank"
+echo "master_ip: $master_ip"
+echo "config: $config"
+echo "output_dir: $output_dir"
+echo "----------------------"
 
 if test -d "$output_dir"; then
     cp $config $output_dir
@@ -58,7 +67,6 @@ NODE_RANK=$node_rank \
 HF_HUB_OFFLINE=0 \
 MASTER_PORT=12348 \
 MASTER_ADDR=$master_ip \
-NCCL_SOCKET_IFNAME=bond1 \
 NCCL_IB_GID_INDEX=3 \
 NCCL_NVLS_ENABLE=0 \
 python3 main.py \
@@ -67,4 +75,3 @@ python3 main.py \
     --config $config \
     --output_dir $output_dir \
     --deepspeed
-