bug fix for Ascend npu

[ash-sigh]

Motivation

Bug fix for Ascend npu

Modifications

1. use torch_npu.npu_scatter_nd_update_ instead of deprecated torch_npu._npu_reshape_and_cache. [Bug] [Ascend] Launching Qwen3-VL-30B-A3B-Instruct got operator error. #11374
2. run image_processor on CPU, because transformers has some limitations on Ascend.
3. fix npu graph index_head_dim AttributeError due to DS-V3.2 changes.Support DeepSeek V3.2 Exp #11061

Accuracy Tests

Server launch script

export HCCL_OP_EXPANSION_MODE="AIV"
#export CPU_AFFINITY_CONF=1,npu0:192-223,npu1:192-223,npu2:128-159,npu3:128-15
export STREAMS_PER_DEVICE=32

python -m sglang.launch_server \
    --model-path /model/Qwen3-VL-30B-A3B-Instruct/ \
    --tp-size 2  --device npu \
    --attention-backend ascend \
    --mm-attention-backend ascend_attn \
    --trust-remote-code

Accuracy test command

python3 -m sglang.test.few_shot_gsm8k --num-questions 200

Accuracy test result

Downloading from https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl to /tmp/test.jsonl
/tmp/test.jsonl: 732kB [00:07, 101kB/s]
100%|███████████████████████████████████████████████████████████████████| 200/200 [00:44<00:00, 4.51it/s]
Accuracy: 0.945
Invalid: 0.000
Latency: 44.533 s
Output throughput: 646.760 token/s

Benchmarking and Profiling

python -m sglang.bench_serving --backend sglang --num-prompt 10 --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

#Input tokens: 1997
#Output tokens: 2798
Starting warmup with 1 sequences...
Warmup completed with 1 sequences. Starting main benchmark run...
100%|███████████████████████████████████████████████| 10/10 [00:22<00:00, 2.20s/it]

============ Serving Benchmark Result ============
Backend: sglang
Traffic request rate: inf
Max request concurrency: not set
Successful requests: 10
Benchmark duration (s): 22.04
Total input tokens: 1997
Total input text tokens: 1997
Total input vision tokens: 0
Total generated tokens: 2798
Total generated tokens (retokenized): 2783
Request throughput (req/s): 0.45
Input token throughput (tok/s): 90.60
Output token throughput (tok/s): 126.94
Total token throughput (tok/s): 217.54
Concurrency: 5.93
----------------End-to-End Latency----------------
Mean E2E Latency (ms): 13079.63
Median E2E Latency (ms): 14950.25
---------------Time to First Token----------------
Mean TTFT (ms): 404.65
Median TTFT (ms): 403.89
P99 TTFT (ms): 450.20
---------------Inter-Token Latency----------------
Mean ITL (ms): 45.50
Median ITL (ms): 46.50
P95 ITL (ms): 50.18
P99 ITL (ms): 52.89
Max ITL (ms): 130.14
==================================================

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.