Issue Title:
[Bug] Single-Node Multi-GPU Training Fails with NCCL Watchdog Timeout despite minimal batch/accumulation (A800 NVLink)

Issue Body:

🐛 Bug Description

I am encountering a persistent NCCL Watchdog Collective Timeout error when training Qwen3-Next-80B-A3B-Instruct with LoRA on a single-node 4x A800 SXM4 server.

Despite aggressively reducing per_device_batch_size to 1 or 2 and gradient_accumulation_steps to 1, the training process hangs at the first few steps (Step 0 or Step 1) and eventually crashes after the default 10-minute timeout (600000ms), throwing Signal 6 (SIGABRT).

Increasing NCCL_TIMEOUT to 3600 or 7200 seems ineffective or is bypassed, as the watchdog still reports a timeout at exactly 600000ms in some runs, or simply hangs indefinitely.

📜 Error Log

[rank1]:[E1227 23:09:04.570] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=56212, OpType=ALLREDUCE, NumelIn=386531328, NumelOut=386531328, Timeout(ms)=600000) ran for 600068 milliseconds before timing out.
[rank1]:[E1227 23:09:04.571] [PG ID 1 PG GUID 1 Rank 1] failure detected by watchdog at work sequence id: 56212 PG status: last enqueued work: 56212, last completed work: 56211
[rank2]:[E1227 23:09:04.716] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=56212, OpType=ALLREDUCE, NumelIn=385990656, NumelOut=385990656, Timeout(ms)=600000) ...
...
terminate called after throwing an instance of 'c10::DistBackendError'
  what():  [PG ID 1 PG GUID 1 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout...

⚙️ Environment

• Hardware: 4x NVIDIA A800-SXM4-80GB (NVLink enabled)
• OS: Linux (Ubuntu 20.04/22.04)
• CUDA: 12.4
• PyTorch Version: (Check with torch.__version__, e.g., 2.4.0+cu121)
• DeepSpeed Version: 0.16.9
• LLaMA-Factory Version: (Latest or commit ID)
• Driver Version: 550.144.03

📋 Reproduction Steps

1. Launch Command:

export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export CUDA_VISIBLE_DEVICES=0,1,2,3
export NCCL_P2P_DISABLE=0  # Tried 0 and 1
export NCCL_IB_DISABLE=0   # Tried 0 and 1
llamafactory-cli webui

2. Training Config (WebUI):

• Model: Qwen3-30B-A3B-Instruct (loaded in bf16)
• Method: LoRA (Rank 32 or 128, trainable params ~100M - 400M)
• Stage: DeepSpeed ZeRO-2
• Per-device batch size: 2
• Gradient accumulation: 8 (also tried 1, still crashes)
• Precision: bf16

🕵️ Troubleshooting Tried

1. Reduced Load: Lowered batch size to 1 and gradient accumulation to 1. Error persists.
2. Reduced Communication: Lowered LoRA Rank to 32. Error persists.
3. Environment Variables: Set NCCL_TIMEOUT=7200 and NCCL_ASYNC_ERROR_HANDLING=1. The process still hangs or crashes.
4. Data: Tried with a single small dataset (1k samples). Error persists.
5. Memory: GPU memory is sufficient (~60GB/80GB used).

❓ Expected Behavior

Training should proceed past Step 0/Step 1 and log loss metrics, utilizing the NVLink interconnect for efficient communication.

Reproduction

Put your message here.

Others

No response