Fix several bugs

Author: ZhangX-21

custom_ops/gpu_ops/append_attn/qwen3_rope.h

@@ -40,6 +40,7 @@ __global__ void GQAVariableLengthRotarySplitKernel_Qwen3(
     const int token_idx = linear_index / offset;
 
     const int ori_bi = batch_id_per_token[token_idx];  // 第几个batch
+    if (ori_bi == -1) continue;  // padded token for static/CUDA-graph shape
 
     int cache_kv_len = seq_lens_decoder[ori_bi];
     // 这里其实是不需要处理的，但是由于FA3的bug，所以必须！

fastdeploy/envs.py

@@ -278,6 +278,8 @@ def _validate_split_kv_size(value: int) -> int:
     # When v1 is enabled, the legacy /clear_load_weight and /update_model_weight
     # will adopt this new communication pattern.
     "FD_ENABLE_V1_UPDATE_WEIGHTS": lambda: bool(int(os.getenv("FD_ENABLE_V1_UPDATE_WEIGHTS", "0"))),
+    # Whether to use GDR CheckpointTransfer for dynamic weight updates.
+    "FD_USE_GDR_CHECKPOINT_TRANSFER": lambda: bool(int(os.getenv("FD_USE_GDR_CHECKPOINT_TRANSFER", "0"))),
     # Whether to save the cache of output token for preempted request to storage.
     "FD_SAVE_OUTPUT_CACHE_FOR_PREEMPTED_REQUEST": lambda: bool(
         int(os.getenv("FD_SAVE_OUTPUT_CACHE_FOR_PREEMPTED_REQUEST", "1"))

fastdeploy/worker/gpu_model_runner.py

@@ -3267,12 +3267,6 @@ def _clear_cache_for_gdr_weight_update(self):
             kv_cache_status.value[0] = KVCacheStatus.CLEARING
         if self.use_cudagraph:
             self.model.clear_graph_opt_backend()
-            if envs.FD_USE_BLOCK_WISE_CUDA_GRAPH:
-                from fastdeploy.model_executor.graph_optimization.cuda_graph_op import (
-                    clear_all_block_wise_graphs,
-                )
-
-                clear_all_block_wise_graphs()
             if (
                 self.speculative_decoding
                 and self.spec_method == SpecMethod.MTP