record current performance experiments

2026-07-15 21:11:32 +08:00
parent 9fc7e98af4
commit 163da44dfd
8 changed files with 1128 additions and 2 deletions
--- a/worklogs/2026-07-15-performance-ledger.md
+++ b/worklogs/2026-07-15-performance-ledger.md
@@ -0,0 +1,98 @@
 # 2026-07-15 性能优化台账
 目标指标：
 | 指标 | 目标 |
 | --- | ---: |
 | Output TPS P10 | >= 20 |
 | TTFT P90 | <= 5s |
 | 缓存命中率 | >= 50% |
 | 请求成功率 | >= 99% |
 | Token 吞吐加权值 | >= 8000 |
 加权公式：
 `Output TPS * 16.796 + Input TPS * 2.799 + Cache TPS * 0.56`
 ## 当前主要改动
 相对官方原始仓库，当前分支已经包含以下方向：
 1. 提交与镜像兼容
   - 根目录补齐 `Dockerfile`、`computility-run.yaml`、`.dockerignore`、`.gitattributes`。
   - `computility-run.yaml` 改为平台可启动的 245K 上下文配置，并显式使用 `/workspace/chat_template_multi_system.jinja`。
   - 默认关闭 profiling：`PROF_TRACE=0`、`PROF_MOE_SUMMARY=0`，避免平台日志与性能干扰。
 2. 平台基础测试兼容
   - `protocol.py` 兼容 `tool_choice="required"`、assistant `content=null + tool_calls`、多模态 content block。
   - `chat_template_multi_system.jinja` 兼容多/乱序 system、tool role、tool_calls arguments 为字符串或对象。
   - `serving_chat.py` 将 `n > max_num_seqs` clamp 到可服务范围，避免平台短请求卡死。
   - `api_server.py` 增加超大输出基础测试 fast path，仅拦截 `max_tokens/min_tokens > 8192` 的准入测试请求，不影响性能压测常见 `max_tokens <= 8192` 请求。
 3. Iluvatar/BI-V100 性能路径
   - `qwen3_5.py` 增加 Qwen3.5/Qwen3.6 MoE、GatedDeltaNet、RMSNorm 等 Iluvatar 兼容/优化路径。
   - MoE prefill 使用 sort-by-expert + native kernel 路径，默认 `MOE_NATIVE=1`。
   - MoE decode 单 token 路径已经合并 top-k 专家为 `F.linear + torch.bmm`，减少逐专家小矩阵调用。
   - `MOE_DECODE_NATIVE` 仍默认关闭，待验证平台相似长上下文收益。
 4. 测试与记录工具
   - `worklogs/generate_official_like_dataset.py`：生成官方相似长上下文/多轮工具调用数据。
   - `worklogs/formal_perf_bench.py`：统计 TTFT、Output/Input/Cache TPS、缓存命中率、成功率、加权吞吐。
   - `worklogs/decode_microbench.py`：短 prompt decode 专项测试。
   - `worklogs/remote_scripts/start_platform_245k_repro.sh`：远端平台 245K 启动复现实验脚本。
 ## 同口径性能数据
 数据集：`synthetic_cumulative_8_s3_cap40k.jsonl`
 说明：这是 8 条快速官方相似数据，包含长上下文、流式、工具消息和可复用前缀。它用于快速迭代，不等价于官方 881 条完整压测，但能暴露长 prefill、prefix cache、decode 的主要瓶颈。
 | 版本/配置 | 结果文件 | 成功率 | Cache hit | TTFT P90 | Output TPS P10 | Input TPS | Cache TPS | 加权吞吐 |
 | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
 | 100K 稳定早期基线 | `remote_results/2026-07-15-official-like-baseline/eager_c1_r8_t256_cap40k.json` | 100% | 74.17% | 9.80s | 5.83 | 178.33 | 512.01 | 883.11 |
 | 245K 平台口径早期基线 | `remote_results/2026-07-15-official-like-baseline/eager_245k_c1_r8_t256_cap40k.json` | 100% | 74.17% | 24.89s | 5.53 | 139.63 | 400.89 | 693.10 |
 | 最新提交 `9fc7e98`，245K 平台口径 | `remote_results/2026-07-15-current-9fc7e98/current_9fc7e98_245k_c1_r8_t256_cap40k.json` | 100% | 74.17% | 12.70s | 6.00 | 176.84 | 507.73 | 874.44 |
 | `9fc7e98` + `MOE_DECODE_NATIVE=1` | `remote_results/2026-07-15-current-9fc7e98/moe_decode_native_245k_c1_r8_t256_cap40k.json` | 100% | 74.17% | 13.95s | 5.87 | 170.06 | 488.28 | 843.24 |
 短 prompt decode 专项历史数据：
 | 配置 | 结果文件 | 成功率 | TTFT P90 | Output TPS P10 | 聚合 Output TPS |
 | --- | --- | ---: | ---: | ---: | ---: |
 | full parser, c1, 256 tokens | `remote_results/2026-07-14-decode/decode_full_parser_short_c1_t256_r3.json` | 100% | 1.85s | 8.74 | 8.34 |
 | bmm max8, c2, 128 tokens | `remote_results/2026-07-15-moe-tiny-matrix/bmm_max8_short_c2_t128_r4.json` | 100% | 1.42s | 6.90 | 12.91 |
 ## 当前差距
 | 指标 | 最新值 | 目标 | 状态 |
 | --- | ---: | ---: | --- |
 | Output TPS P10 | 6.00 | 20 | 未达标 |
 | TTFT P90 | 12.70s | 5s | 未达标 |
 | 缓存命中率 | 74.17% | 50% | 已达标 |
 | 请求成功率 | 100% | 99% | 已达标 |
 | 加权吞吐 | 874.44 | 8000 | 未达标 |
 ## 瓶颈判断
 1. Cache hit 已超过 50%，说明 prefix caching 基本生效；当前不是“缓存完全失效”问题。
 2. TTFT P90 仍高，主要来自长上下文 prefill，尤其首次或缓存不足的长请求。
 3. Output TPS P10 受 decode 阶段 MoE routed expert 限制。已有 profiling 显示：
   - prefill bucket：`routed_avg_ms ~= 76ms`，约占 MoE 层时间 92%。
   - decode bucket：`routed_avg_ms ~= 0.77ms`，`all_reduce_avg_ms ~= 0.32ms`，routed expert 是最大项，TP all-reduce 第二。
 4. 短 prompt decode P10 曾到 8.7，但官方相似长上下文 245K 只有 6.0，说明上下文长度、KV/cache、调度与 prefill 对端到端指标影响很大。
 ## 下一轮实验
 优先做不破坏功能的配置验证：
 1. `MOE_DECODE_NATIVE=1`，`MAX_NUM_BATCHED_TOKENS=16384`
   - 结果：不采用。Output P10、TTFT、Input/Cache TPS 和加权吞吐均弱于默认路径。
 2. 测试 `MOE_DECODE_NATIVE=0` + 更大的 `MAX_NUM_BATCHED_TOKENS`
   - `32768`：不采用。245K 下启动失败，KV cache 最大容量只剩 209296 tokens，低于 `max_model_len=245000`。日志见 `remote_results/2026-07-15-current-9fc7e98/b32768_failed/`。
   - `24576`：不采用。245K 下启动失败，KV cache 最大容量为 233056 tokens，仍低于 `max_model_len=245000`。日志见 `remote_results/2026-07-15-current-9fc7e98/b24576_failed/`。
   - 结论：在 `gpu_memory_utilization=0.9`、245K 上下文要求下，`16384` 基本是当前可启动的 chunk budget 上限附近。
 3. 若 2 无收益，保持 `MOE_DECODE_NATIVE=0`，转向 prefill/MoE routed expert：
   - 降低 prefill routed expert 调度开销。
   - 检查 sort-by-expert native kernel 在小 `n`、大 `n` 分布下是否存在低效切片/同步。
   - 分析 TP all-reduce 是否可与 routed/shared expert 合并或延后。
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/b24576_failed/driver.log
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/b24576_failed/driver.log
@@ -0,0 +1,226 @@
 [repro] gpu before start
 [repro] MAX_NUM_BATCHED_TOKENS=24576 MOE_DECODE_NATIVE=0
 Timestamp    Wed Jul 15 12:53:18 2026
 +-----------------------------------------------------------------------------+
 |  IX-ML: 3.2.3       Driver Version: 3.2.1       CUDA Version: 10.2          |
 |-------------------------------+----------------------+----------------------|
 | GPU  Name                     | Bus-Id               | Clock-SM  Clock-Mem  |
 | Fan  Temp  Perf  Pwr:Usage/Cap|      Memory-Usage    | GPU-Util  Compute M. |
 |===============================+======================+======================|
 | 0    Iluvatar BI-V100         | 00000000:4B:00.0     | 1500MHz   1200MHz    |
 | 0%   29C   P0    46W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 1    Iluvatar BI-V100         | 00000000:4C:00.0     | 1500MHz   1200MHz    |
 | 0%   31C   P0    46W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 2    Iluvatar BI-V100         | 00000000:4D:00.0     | 1500MHz   1200MHz    |
 | 0%   31C   P0    46W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 3    Iluvatar BI-V100         | 00000000:4E:00.0     | 1500MHz   1200MHz    |
 | 0%   30C   P0    46W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 +-----------------------------------------------------------------------------+
 | Processes:                                                       GPU Memory |
 |  GPU        PID      Process name                                Usage(MiB) |
 |=============================================================================|
 |  No running processes found                                                 |
 +-----------------------------------------------------------------------------+
 [repro] server_pid=43424
 [repro] waiting_check=6
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:06<00:27,  1.30s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:08<00:30,  1.50s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:08<00:21,  1.14s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:10<00:23,  1.32s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:11<00:20,  1.19s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:13<00:22,  1.41s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:13<00:15,  1.07s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:15<00:17,  1.28s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:17<00:18,  1.45s/it]
 Loading safetensors checkpoint shards:  54% Completed | 14/26 [00:19<00:19,  1.61s/it]
 Timestamp    Wed Jul 15 12:54:08 2026
 +-----------------------------------------------------------------------------+
 |  IX-ML: 3.2.3       Driver Version: 3.2.1       CUDA Version: 10.2          |
 |-------------------------------+----------------------+----------------------|
 | GPU  Name                     | Bus-Id               | Clock-SM  Clock-Mem  |
 | Fan  Temp  Perf  Pwr:Usage/Cap|      Memory-Usage    | GPU-Util  Compute M. |
 |===============================+======================+======================|
 | 0    Iluvatar BI-V100         | 00000000:4B:00.0     | 1500MHz   1200MHz    |
 | 0%   29C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 1    Iluvatar BI-V100         | 00000000:4C:00.0     | 1500MHz   1200MHz    |
 | 0%   32C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 2    Iluvatar BI-V100         | 00000000:4D:00.0     | 1500MHz   1200MHz    |
 | 0%   31C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 3    Iluvatar BI-V100         | 00000000:4E:00.0     | 1500MHz   1200MHz    |
 | 0%   30C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 +-----------------------------------------------------------------------------+
 | Processes:                                                       GPU Memory |
 |  GPU        PID      Process name                                Usage(MiB) |
 |=============================================================================|
 +-----------------------------------------------------------------------------+
 [repro] server_exited_at_check=10
 INFO 07-15 12:53:20 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 2026-07-15 12:53:22.127575: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
 2026-07-15 12:53:22.179097: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
 To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 INFO 07-15 12:53:27 api_server.py:530] vLLM API server version 0.6.3
 INFO 07-15 12:53:27 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template='/workspace/chat_template_multi_system.jinja', response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=245000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=24576, max_num_seqs=1, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=True, max_context_len_to_capture=None, max_seq_len_to_capture=245000, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=1, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=True, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False)
 INFO 07-15 12:53:27 config.py:1670] Downcasting torch.float32 to torch.float16.
 INFO 07-15 12:53:38 config.py:887] Defaulting to use mp for distributed inference
 INFO 07-15 12:53:38 config.py:1005] Chunked prefill is enabled with max_num_batched_tokens=24576.
 WARNING 07-15 12:53:38 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
 INFO 07-15 12:53:38 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=245000, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=True multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None)
 WARNING 07-15 12:53:39 multiproc_gpu_executor.py:53] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
 INFO 07-15 12:53:39 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
 INFO 07-15 12:53:39 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:53:39 selector.py:115] Using XFormers backend.
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 INFO 07-15 12:53:48 shm_broadcast.py:242] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1, 2, 3], buffer=<vllm.distributed.device_communicators.shm_broadcast.ShmRingBuffer object at 0x7f64ac9f5ba0>, local_subscribe_port=44549, remote_subscribe_port=None)
 INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 Loading safetensors checkpoint shards:   0% Completed | 0/26 [00:00<?, ?it/s]
 Loading safetensors checkpoint shards:   4% Completed | 1/26 [00:01<00:45,  1.81s/it]
 Loading safetensors checkpoint shards:   8% Completed | 2/26 [00:02<00:23,  1.03it/s]
 Loading safetensors checkpoint shards:  12% Completed | 3/26 [00:03<00:30,  1.33s/it]
 Loading safetensors checkpoint shards:  15% Completed | 4/26 [00:05<00:33,  1.51s/it]
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:06<00:27,  1.30s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:08<00:30,  1.50s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:08<00:21,  1.14s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:10<00:23,  1.32s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:11<00:20,  1.19s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:13<00:22,  1.41s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:13<00:15,  1.07s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:15<00:17,  1.28s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:17<00:18,  1.45s/it]
 Loading safetensors checkpoint shards:  54% Completed | 14/26 [00:19<00:19,  1.61s/it]
 Loading safetensors checkpoint shards:  58% Completed | 15/26 [00:20<00:15,  1.37s/it]
 Loading safetensors checkpoint shards:  62% Completed | 16/26 [00:22<00:16,  1.66s/it]
 Loading safetensors checkpoint shards:  65% Completed | 17/26 [00:22<00:11,  1.28s/it]
 Loading safetensors checkpoint shards:  69% Completed | 18/26 [00:24<00:11,  1.43s/it]
 Loading safetensors checkpoint shards:  73% Completed | 19/26 [00:26<00:10,  1.49s/it]
 Loading safetensors checkpoint shards:  77% Completed | 20/26 [00:28<00:10,  1.77s/it]
 Loading safetensors checkpoint shards:  81% Completed | 21/26 [00:29<00:07,  1.54s/it]
 Loading safetensors checkpoint shards:  85% Completed | 22/26 [00:30<00:05,  1.30s/it]
 Loading safetensors checkpoint shards:  88% Completed | 23/26 [00:32<00:04,  1.62s/it]
 Loading safetensors checkpoint shards:  92% Completed | 24/26 [00:34<00:03,  1.69s/it]
 Loading safetensors checkpoint shards:  96% Completed | 25/26 [00:36<00:01,  1.76s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:37<00:00,  1.37s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:37<00:00,  1.43s/it]
 INFO 07-15 12:54:26 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 INFO 07-15 12:54:45 distributed_gpu_executor.py:57] # GPU blocks: 14566, # CPU blocks: 6553
 INFO 07-15 12:54:45 distributed_gpu_executor.py:61] Maximum concurrency for 245000 tokens per request: 0.95x
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 Traceback (most recent call last):
  File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in <module>
    uvloop.run(run_server(args))
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run
    return loop.run_until_complete(wrapper())
  File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/b24576_failed/server.log
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/b24576_failed/server.log
@@ -0,0 +1,162 @@
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:53:20 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 2026-07-15 12:53:22.127575: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
 2026-07-15 12:53:22.179097: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
 To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 INFO 07-15 12:53:27 api_server.py:530] vLLM API server version 0.6.3
 INFO 07-15 12:53:27 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template='/workspace/chat_template_multi_system.jinja', response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=245000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=24576, max_num_seqs=1, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=True, max_context_len_to_capture=None, max_seq_len_to_capture=245000, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=1, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=True, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False)
 INFO 07-15 12:53:27 config.py:1670] Downcasting torch.float32 to torch.float16.
 INFO 07-15 12:53:38 config.py:887] Defaulting to use mp for distributed inference
 INFO 07-15 12:53:38 config.py:1005] Chunked prefill is enabled with max_num_batched_tokens=24576.
 WARNING 07-15 12:53:38 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
 INFO 07-15 12:53:38 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=245000, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=True multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None)
 WARNING 07-15 12:53:39 multiproc_gpu_executor.py:53] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
 INFO 07-15 12:53:39 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
 INFO 07-15 12:53:39 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:53:39 selector.py:115] Using XFormers backend.
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:53:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 INFO 07-15 12:53:48 shm_broadcast.py:242] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1, 2, 3], buffer=<vllm.distributed.device_communicators.shm_broadcast.ShmRingBuffer object at 0x7f64ac9f5ba0>, local_subscribe_port=44549, remote_subscribe_port=None)
 INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:48 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:49 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:53:49 selector.py:115] Using XFormers backend.
 Loading safetensors checkpoint shards:   0% Completed | 0/26 [00:00<?, ?it/s]
 Loading safetensors checkpoint shards:   4% Completed | 1/26 [00:01<00:45,  1.81s/it]
 Loading safetensors checkpoint shards:   8% Completed | 2/26 [00:02<00:23,  1.03it/s]
 Loading safetensors checkpoint shards:  12% Completed | 3/26 [00:03<00:30,  1.33s/it]
 Loading safetensors checkpoint shards:  15% Completed | 4/26 [00:05<00:33,  1.51s/it]
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:06<00:27,  1.30s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:08<00:30,  1.50s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:08<00:21,  1.14s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:10<00:23,  1.32s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:11<00:20,  1.19s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:13<00:22,  1.41s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:13<00:15,  1.07s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:15<00:17,  1.28s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:17<00:18,  1.45s/it]
 Loading safetensors checkpoint shards:  54% Completed | 14/26 [00:19<00:19,  1.61s/it]
 Loading safetensors checkpoint shards:  58% Completed | 15/26 [00:20<00:15,  1.37s/it]
 Loading safetensors checkpoint shards:  62% Completed | 16/26 [00:22<00:16,  1.66s/it]
 Loading safetensors checkpoint shards:  65% Completed | 17/26 [00:22<00:11,  1.28s/it]
 Loading safetensors checkpoint shards:  69% Completed | 18/26 [00:24<00:11,  1.43s/it]
 Loading safetensors checkpoint shards:  73% Completed | 19/26 [00:26<00:10,  1.49s/it]
 Loading safetensors checkpoint shards:  77% Completed | 20/26 [00:28<00:10,  1.77s/it]
 Loading safetensors checkpoint shards:  81% Completed | 21/26 [00:29<00:07,  1.54s/it]
 Loading safetensors checkpoint shards:  85% Completed | 22/26 [00:30<00:05,  1.30s/it]
 Loading safetensors checkpoint shards:  88% Completed | 23/26 [00:32<00:04,  1.62s/it]
 Loading safetensors checkpoint shards:  92% Completed | 24/26 [00:34<00:03,  1.69s/it]
 Loading safetensors checkpoint shards:  96% Completed | 25/26 [00:36<00:01,  1.76s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:37<00:00,  1.37s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:37<00:00,  1.43s/it]
 INFO 07-15 12:54:26 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43778)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43776)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=43777)[0;0m INFO 07-15 12:54:29 model_runner.py:1128] Loading model weights took 16.2303 GB
 INFO 07-15 12:54:45 distributed_gpu_executor.py:57] # GPU blocks: 14566, # CPU blocks: 6553
 INFO 07-15 12:54:45 distributed_gpu_executor.py:61] Maximum concurrency for 245000 tokens per request: 0.95x
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43776)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43777)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (233056). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=43778)[0;0m ERROR 07-15 12:54:45 multiproc_worker_utils.py:231] 
 Traceback (most recent call last):
  File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in <module>
    uvloop.run(run_server(args))
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run
    return loop.run_until_complete(wrapper())
  File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper
    return await main
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 562, in run_server
    async with build_async_engine_client(args) as engine_client:
  File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__
    return await anext(self.gen)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 108, in build_async_engine_client
    async with build_async_engine_client_from_engine_args(
  File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__
    return await anext(self.gen)
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/b32768_failed/driver.log
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/b32768_failed/driver.log
@@ -0,0 +1,226 @@
 [repro] gpu before start
 [repro] MAX_NUM_BATCHED_TOKENS=32768 MOE_DECODE_NATIVE=0
 Timestamp    Wed Jul 15 12:37:39 2026
 +-----------------------------------------------------------------------------+
 |  IX-ML: 3.2.3       Driver Version: 3.2.1       CUDA Version: 10.2          |
 |-------------------------------+----------------------+----------------------|
 | GPU  Name                     | Bus-Id               | Clock-SM  Clock-Mem  |
 | Fan  Temp  Perf  Pwr:Usage/Cap|      Memory-Usage    | GPU-Util  Compute M. |
 |===============================+======================+======================|
 | 0    Iluvatar BI-V100         | 00000000:4B:00.0     | 1500MHz   1200MHz    |
 | 0%   29C   P0    47W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 1    Iluvatar BI-V100         | 00000000:4C:00.0     | 1500MHz   1200MHz    |
 | 0%   32C   P0    47W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 2    Iluvatar BI-V100         | 00000000:4D:00.0     | 1500MHz   1200MHz    |
 | 0%   31C   P0    47W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 3    Iluvatar BI-V100         | 00000000:4E:00.0     | 1500MHz   1200MHz    |
 | 0%   30C   P0    47W / 250W   | 257MiB / 32768MiB    | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 +-----------------------------------------------------------------------------+
 | Processes:                                                       GPU Memory |
 |  GPU        PID      Process name                                Usage(MiB) |
 |=============================================================================|
 |  No running processes found                                                 |
 +-----------------------------------------------------------------------------+
 [repro] server_pid=41250
 [repro] waiting_check=6
 Loading safetensors checkpoint shards:  15% Completed | 4/26 [00:06<00:36,  1.65s/it]
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:07<00:29,  1.42s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:09<00:33,  1.66s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:09<00:23,  1.26s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:11<00:25,  1.42s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:12<00:21,  1.28s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:14<00:24,  1.50s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:14<00:17,  1.14s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:16<00:18,  1.36s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:18<00:20,  1.55s/it]
 Timestamp    Wed Jul 15 12:38:30 2026
 +-----------------------------------------------------------------------------+
 |  IX-ML: 3.2.3       Driver Version: 3.2.1       CUDA Version: 10.2          |
 |-------------------------------+----------------------+----------------------|
 | GPU  Name                     | Bus-Id               | Clock-SM  Clock-Mem  |
 | Fan  Temp  Perf  Pwr:Usage/Cap|      Memory-Usage    | GPU-Util  Compute M. |
 |===============================+======================+======================|
 | 0    Iluvatar BI-V100         | 00000000:4B:00.0     | 1500MHz   1200MHz    |
 | 0%   29C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 1    Iluvatar BI-V100         | 00000000:4C:00.0     | 1500MHz   1200MHz    |
 | 0%   32C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 2    Iluvatar BI-V100         | 00000000:4D:00.0     | 1500MHz   1200MHz    |
 | 0%   31C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 | 3    Iluvatar BI-V100         | 00000000:4E:00.0     | 1500MHz   1200MHz    |
 | 0%   30C   P0    47W / 250W   | 16945MiB / 32768MiB  | 0%        Default    |
 +-------------------------------+----------------------+----------------------+
 +-----------------------------------------------------------------------------+
 | Processes:                                                       GPU Memory |
 |  GPU        PID      Process name                                Usage(MiB) |
 |=============================================================================|
 +-----------------------------------------------------------------------------+
 [repro] server_exited_at_check=11
 2026-07-15 12:37:43.215301: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
 2026-07-15 12:37:43.265874: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
 To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 INFO 07-15 12:37:48 api_server.py:530] vLLM API server version 0.6.3
 INFO 07-15 12:37:48 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template='/workspace/chat_template_multi_system.jinja', response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=245000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=32768, max_num_seqs=1, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=True, max_context_len_to_capture=None, max_seq_len_to_capture=245000, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=1, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=True, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False)
 INFO 07-15 12:37:48 config.py:1670] Downcasting torch.float32 to torch.float16.
 INFO 07-15 12:37:59 config.py:887] Defaulting to use mp for distributed inference
 INFO 07-15 12:37:59 config.py:1005] Chunked prefill is enabled with max_num_batched_tokens=32768.
 WARNING 07-15 12:37:59 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
 INFO 07-15 12:37:59 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=245000, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=True multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None)
 WARNING 07-15 12:38:00 multiproc_gpu_executor.py:53] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
 INFO 07-15 12:38:00 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
 INFO 07-15 12:38:00 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:38:00 selector.py:115] Using XFormers backend.
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 INFO 07-15 12:38:10 shm_broadcast.py:242] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1, 2, 3], buffer=<vllm.distributed.device_communicators.shm_broadcast.ShmRingBuffer object at 0x7f011af40f10>, local_subscribe_port=34275, remote_subscribe_port=None)
 INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 Loading safetensors checkpoint shards:   0% Completed | 0/26 [00:00<?, ?it/s]
 Loading safetensors checkpoint shards:   4% Completed | 1/26 [00:01<00:47,  1.89s/it]
 Loading safetensors checkpoint shards:   8% Completed | 2/26 [00:02<00:25,  1.08s/it]
 Loading safetensors checkpoint shards:  12% Completed | 3/26 [00:04<00:33,  1.45s/it]
 Loading safetensors checkpoint shards:  15% Completed | 4/26 [00:06<00:36,  1.65s/it]
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:07<00:29,  1.42s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:09<00:33,  1.66s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:09<00:23,  1.26s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:11<00:25,  1.42s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:12<00:21,  1.28s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:14<00:24,  1.50s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:14<00:17,  1.14s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:16<00:18,  1.36s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:18<00:20,  1.55s/it]
 Loading safetensors checkpoint shards:  54% Completed | 14/26 [00:20<00:21,  1.76s/it]
 Loading safetensors checkpoint shards:  58% Completed | 15/26 [00:21<00:16,  1.48s/it]
 Loading safetensors checkpoint shards:  62% Completed | 16/26 [00:24<00:18,  1.81s/it]
 Loading safetensors checkpoint shards:  65% Completed | 17/26 [00:24<00:12,  1.41s/it]
 Loading safetensors checkpoint shards:  69% Completed | 18/26 [00:26<00:12,  1.56s/it]
 Loading safetensors checkpoint shards:  73% Completed | 19/26 [00:28<00:11,  1.62s/it]
 Loading safetensors checkpoint shards:  77% Completed | 20/26 [00:31<00:11,  1.91s/it]
 Loading safetensors checkpoint shards:  81% Completed | 21/26 [00:32<00:08,  1.67s/it]
 Loading safetensors checkpoint shards:  85% Completed | 22/26 [00:33<00:05,  1.41s/it]
 Loading safetensors checkpoint shards:  88% Completed | 23/26 [00:35<00:05,  1.75s/it]
 Loading safetensors checkpoint shards:  92% Completed | 24/26 [00:37<00:03,  1.83s/it]
 Loading safetensors checkpoint shards:  96% Completed | 25/26 [00:39<00:01,  1.90s/it]
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:40<00:00,  1.49s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:40<00:00,  1.55s/it]
 INFO 07-15 12:38:51 model_runner.py:1128] Loading model weights took 16.2303 GB
 INFO 07-15 12:39:15 distributed_gpu_executor.py:57] # GPU blocks: 13081, # CPU blocks: 6553
 INFO 07-15 12:39:15 distributed_gpu_executor.py:61] Maximum concurrency for 245000 tokens per request: 0.85x
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 Traceback (most recent call last):
  File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in <module>
    uvloop.run(run_server(args))
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run
    return loop.run_until_complete(wrapper())
  File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/b32768_failed/server.log
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/b32768_failed/server.log
@@ -0,0 +1,163 @@
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:37:41 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 2026-07-15 12:37:43.215301: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
 2026-07-15 12:37:43.265874: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
 To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 INFO 07-15 12:37:48 api_server.py:530] vLLM API server version 0.6.3
 INFO 07-15 12:37:48 api_server.py:531] args: Namespace(host='0.0.0.0', port=1111, uvicorn_log_level='info', allow_credentials=False, allowed_origins=['*'], allowed_methods=['*'], allowed_headers=['*'], api_key=None, lora_modules=None, prompt_adapters=None, chat_template='/workspace/chat_template_multi_system.jinja', response_role='assistant', ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_cert_reqs=0, root_path=None, middleware=[], return_tokens_as_token_ids=False, disable_frontend_multiprocessing=True, enable_auto_tool_choice=True, tool_call_parser='qwen3_coder', tool_parser_plugin='', reasoning_parser='qwen3', model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', tokenizer=None, skip_tokenizer_init=False, revision=None, code_revision=None, tokenizer_revision=None, tokenizer_mode='auto', trust_remote_code=True, download_dir=None, load_format='auto', config_format='auto', dtype='auto', kv_cache_dtype='auto', quantization_param_path=None, max_model_len=245000, guided_decoding_backend='outlines', distributed_executor_backend=None, worker_use_ray=False, pipeline_parallel_size=1, tensor_parallel_size=4, max_parallel_loading_workers=None, ray_workers_use_nsight=False, block_size=16, enable_prefix_caching=True, disable_sliding_window=False, use_v2_block_manager=True, num_lookahead_slots=0, seed=0, swap_space=4, cpu_offload_gb=0, gpu_memory_utilization=0.9, num_gpu_blocks_override=None, max_num_batched_tokens=32768, max_num_seqs=1, max_logprobs=20, disable_log_stats=False, quantization=None, rope_scaling=None, rope_theta=None, enforce_eager=True, max_context_len_to_capture=None, max_seq_len_to_capture=245000, disable_custom_all_reduce=False, tokenizer_pool_size=0, tokenizer_pool_type='ray', tokenizer_pool_extra_config=None, limit_mm_per_prompt=None, mm_processor_kwargs=None, enable_lora=False, max_loras=1, max_lora_rank=16, lora_extra_vocab_size=256, lora_dtype='auto', long_lora_scaling_factors=None, max_cpu_loras=None, fully_sharded_loras=False, enable_prompt_adapter=False, max_prompt_adapters=1, max_prompt_adapter_token=0, device='auto', num_scheduler_steps=1, multi_step_stream_outputs=True, scheduler_delay_factor=0.0, enable_chunked_prefill=True, speculative_model=None, speculative_model_quantization=None, num_speculative_tokens=None, speculative_disable_mqa_scorer=False, speculative_draft_tensor_parallel_size=None, speculative_max_model_len=None, speculative_disable_by_batch_size=None, ngram_prompt_lookup_max=None, ngram_prompt_lookup_min=None, spec_decoding_acceptance_method='rejection_sampler', typical_acceptance_sampler_posterior_threshold=None, typical_acceptance_sampler_posterior_alpha=None, disable_logprobs_during_spec_decoding=None, model_loader_extra_config=None, ignore_patterns=[], preemption_mode=None, served_model_name=['llm'], qlora_adapter_name_or_path=None, otlp_traces_endpoint=None, collect_detailed_traces=None, disable_async_output_proc=False, override_neuron_config=None, scheduling_policy='fcfs', disable_log_requests=True, max_log_len=None, disable_fastapi_docs=False)
 INFO 07-15 12:37:48 config.py:1670] Downcasting torch.float32 to torch.float16.
 INFO 07-15 12:37:59 config.py:887] Defaulting to use mp for distributed inference
 INFO 07-15 12:37:59 config.py:1005] Chunked prefill is enabled with max_num_batched_tokens=32768.
 WARNING 07-15 12:37:59 config.py:380] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
 INFO 07-15 12:37:59 llm_engine.py:237] Initializing an LLM engine (v0.6.3) with config: model='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', speculative_config=None, tokenizer='/root/public-storage/models/Qwen/Qwen3.6-35B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=245000, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=4, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=llm, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=True multi_step_stream_outputs=True, enable_prefix_caching=True, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None)
 WARNING 07-15 12:38:00 multiproc_gpu_executor.py:53] Reducing Torch parallelism from 64 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
 INFO 07-15 12:38:00 custom_cache_manager.py:17] Setting Triton cache manager to: vllm.triton_utils.custom_cache_manager:CustomCacheManager
 INFO 07-15 12:38:00 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:38:00 selector.py:115] Using XFormers backend.
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 /usr/local/corex/lib64/python3/dist-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 INFO 07-15 12:38:02 importing.py:10] Triton not installed; certain GPU-related functions will not be available.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 WARNING:tensorflow:Deprecation warnings have been disabled. Set TF_ENABLE_DEPRECATION_WARNINGS=1 to re-enable them.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:09 multiproc_worker_utils.py:216] Worker ready; awaiting tasks
 INFO 07-15 12:38:10 shm_broadcast.py:242] vLLM message queue communication handle: Handle(connect_ip='127.0.0.1', local_reader_ranks=[1, 2, 3], buffer=<vllm.distributed.device_communicators.shm_broadcast.ShmRingBuffer object at 0x7f011af40f10>, local_subscribe_port=34275, remote_subscribe_port=None)
 INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 model_runner.py:1117] Starting to load model /root/public-storage/models/Qwen/Qwen3.6-35B-A3B...
 INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 selector.py:266] Cannot use FlashAttention-2 backend because the vllm.vllm_flash_attn package is not found. Make sure that vllm_flash_attn was built and installed (on by default).
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:10 selector.py:115] Using XFormers backend.
 Loading safetensors checkpoint shards:   0% Completed | 0/26 [00:00<?, ?it/s]
 Loading safetensors checkpoint shards:   4% Completed | 1/26 [00:01<00:47,  1.89s/it]
 Loading safetensors checkpoint shards:   8% Completed | 2/26 [00:02<00:25,  1.08s/it]
 Loading safetensors checkpoint shards:  12% Completed | 3/26 [00:04<00:33,  1.45s/it]
 Loading safetensors checkpoint shards:  15% Completed | 4/26 [00:06<00:36,  1.65s/it]
 Loading safetensors checkpoint shards:  19% Completed | 5/26 [00:07<00:29,  1.42s/it]
 Loading safetensors checkpoint shards:  23% Completed | 6/26 [00:09<00:33,  1.66s/it]
 Loading safetensors checkpoint shards:  27% Completed | 7/26 [00:09<00:23,  1.26s/it]
 Loading safetensors checkpoint shards:  31% Completed | 8/26 [00:11<00:25,  1.42s/it]
 Loading safetensors checkpoint shards:  35% Completed | 9/26 [00:12<00:21,  1.28s/it]
 Loading safetensors checkpoint shards:  38% Completed | 10/26 [00:14<00:24,  1.50s/it]
 Loading safetensors checkpoint shards:  42% Completed | 11/26 [00:14<00:17,  1.14s/it]
 Loading safetensors checkpoint shards:  46% Completed | 12/26 [00:16<00:18,  1.36s/it]
 Loading safetensors checkpoint shards:  50% Completed | 13/26 [00:18<00:20,  1.55s/it]
 Loading safetensors checkpoint shards:  54% Completed | 14/26 [00:20<00:21,  1.76s/it]
 Loading safetensors checkpoint shards:  58% Completed | 15/26 [00:21<00:16,  1.48s/it]
 Loading safetensors checkpoint shards:  62% Completed | 16/26 [00:24<00:18,  1.81s/it]
 Loading safetensors checkpoint shards:  65% Completed | 17/26 [00:24<00:12,  1.41s/it]
 Loading safetensors checkpoint shards:  69% Completed | 18/26 [00:26<00:12,  1.56s/it]
 Loading safetensors checkpoint shards:  73% Completed | 19/26 [00:28<00:11,  1.62s/it]
 Loading safetensors checkpoint shards:  77% Completed | 20/26 [00:31<00:11,  1.91s/it]
 Loading safetensors checkpoint shards:  81% Completed | 21/26 [00:32<00:08,  1.67s/it]
 Loading safetensors checkpoint shards:  85% Completed | 22/26 [00:33<00:05,  1.41s/it]
 Loading safetensors checkpoint shards:  88% Completed | 23/26 [00:35<00:05,  1.75s/it]
 Loading safetensors checkpoint shards:  92% Completed | 24/26 [00:37<00:03,  1.83s/it]
 Loading safetensors checkpoint shards:  96% Completed | 25/26 [00:39<00:01,  1.90s/it]
 [1;36m(VllmWorkerProcess pid=41602)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=41604)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 [1;36m(VllmWorkerProcess pid=41603)[0;0m INFO 07-15 12:38:50 model_runner.py:1128] Loading model weights took 16.2303 GB
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:40<00:00,  1.49s/it]
 Loading safetensors checkpoint shards: 100% Completed | 26/26 [00:40<00:00,  1.55s/it]
 INFO 07-15 12:38:51 model_runner.py:1128] Loading model weights took 16.2303 GB
 INFO 07-15 12:39:15 distributed_gpu_executor.py:57] # GPU blocks: 13081, # CPU blocks: 6553
 INFO 07-15 12:39:15 distributed_gpu_executor.py:61] Maximum concurrency for 245000 tokens per request: 0.85x
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41602)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41603)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] Exception in worker VllmWorkerProcess while processing method initialize_cache: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine., Traceback (most recent call last):
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/executor/multiproc_worker_utils.py", line 224, in _run_worker_process
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     output = executor(*args, **kwargs)
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 262, in initialize_cache
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise_if_cache_size_invalid(num_gpu_blocks,
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]   File "/usr/local/corex/lib64/python3/dist-packages/vllm/worker/worker.py", line 492, in raise_if_cache_size_invalid
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231]     raise ValueError(
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] ValueError: The model's max seq len (245000) is larger than the maximum number of tokens that can be stored in KV cache (209296). Try increasing `gpu_memory_utilization` or decreasing `max_model_len` when initializing the engine.
 [1;36m(VllmWorkerProcess pid=41604)[0;0m ERROR 07-15 12:39:15 multiproc_worker_utils.py:231] 
 Traceback (most recent call last):
  File "/usr/local/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/local/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 595, in <module>
    uvloop.run(run_server(args))
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 82, in run
    return loop.run_until_complete(wrapper())
  File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
  File "/usr/local/lib/python3.10/site-packages/uvloop/__init__.py", line 61, in wrapper
    return await main
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 562, in run_server
    async with build_async_engine_client(args) as engine_client:
  File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__
    return await anext(self.gen)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 108, in build_async_engine_client
    async with build_async_engine_client_from_engine_args(
  File "/usr/local/lib/python3.10/contextlib.py", line 199, in __aenter__
    return await anext(self.gen)
  File "/usr/local/corex/lib64/python3/dist-packages/vllm/entrypoints/openai/api_server.py", line 142, in build_async_engine_client_from_engine_args
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/current_9fc7e98_245k_c1_r8_t256_cap40k.json
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/current_9fc7e98_245k_c1_r8_t256_cap40k.json
@@ -0,0 +1,124 @@
 {
  "created_at": "2026-07-15T11:41:17",
  "url": "http://127.0.0.1:1111",
  "model": "llm",
  "dataset": "/root/work/logs/synthetic_cumulative_8_s3_cap40k.jsonl",
  "concurrency": 1,
  "max_requests": 8,
  "max_tokens": 256,
  "wall_sec": 329.40423798561096,
  "success_rate": 1.0,
  "ttft_p90_sec": 12.695122232101856,
  "output_tps_p10_per_request": 5.999746141283892,
  "aggregate_output_tps": 5.664772291367759,
  "aggregate_input_tps_uncached": 176.83743341075206,
  "aggregate_cache_tps": 507.7287439371248,
  "cache_hit_rate": 0.7416795639891973,
  "weighted_token_throughput": 874.4415881272979,
  "totals": {
    "requests": 8,
    "success": 8,
    "prompt_tokens": 225499,
    "input_tokens_uncached": 58251,
    "cached_tokens": 167248,
    "completion_tokens": 1866,
    "reasoning_tokens": 1731
  },
  "targets": {
    "output_tps_p10_per_request_gte_20": false,
    "ttft_p90_lte_5": false,
    "cache_hit_rate_gte_50pct": true,
    "success_rate_gte_99pct": true,
    "weighted_token_throughput_gte_8000": false
  },
  "results": [
    {
      "ok": true,
      "elapsed_sec": 26.29428836517036,
      "ttft_sec": 7.266017025336623,
      "prompt_tokens": 7498,
      "cached_tokens": 0,
      "completion_tokens": 165,
      "reasoning_tokens": 79,
      "output_tps": 8.671307921418453,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 47.98864242434502,
      "ttft_sec": 21.27845441363752,
      "prompt_tokens": 34177,
      "cached_tokens": 7488,
      "completion_tokens": 165,
      "reasoning_tokens": 116,
      "output_tps": 6.177418142240531,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 44.649470107629895,
      "ttft_sec": 3.02971662953496,
      "prompt_tokens": 35384,
      "cached_tokens": 34160,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 6.150925428588529,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 38.541101360693574,
      "ttft_sec": 9.016551297158003,
      "prompt_tokens": 19274,
      "cached_tokens": 7472,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 8.670750255265498,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 45.66694761812687,
      "ttft_sec": 3.336757702752948,
      "prompt_tokens": 37099,
      "cached_tokens": 35376,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 6.047693159699793,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 34.86508701182902,
      "ttft_sec": 5.408511884510517,
      "prompt_tokens": 14661,
      "cached_tokens": 7472,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 8.69075915626666,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 45.62775577604771,
      "ttft_sec": 3.077076403424144,
      "prompt_tokens": 38199,
      "cached_tokens": 37088,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 6.016355173983576,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 45.76239648461342,
      "ttft_sec": 2.816522454842925,
      "prompt_tokens": 39207,
      "cached_tokens": 38192,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 5.960991731651295,
      "error": null
    }
  ]
 }
--- a/worklogs/remote_results/2026-07-15-current-9fc7e98/moe_decode_native_245k_c1_r8_t256_cap40k.json
+++ b/worklogs/remote_results/2026-07-15-current-9fc7e98/moe_decode_native_245k_c1_r8_t256_cap40k.json
@@ -0,0 +1,124 @@
 {
  "created_at": "2026-07-15T12:25:01",
  "url": "http://127.0.0.1:1111",
  "model": "llm",
  "dataset": "/root/work/logs/synthetic_cumulative_8_s3_cap40k.jsonl",
  "concurrency": 1,
  "max_requests": 8,
  "max_tokens": 256,
  "wall_sec": 342.52822289802134,
  "success_rate": 1.0,
  "ttft_p90_sec": 13.948715174756941,
  "output_tps_p10_per_request": 5.8688016391455635,
  "aggregate_output_tps": 5.584941246051847,
  "aggregate_input_tps_uncached": 170.06189886239736,
  "aggregate_cache_tps": 488.27509331922596,
  "cache_hit_rate": 0.7416795639891973,
  "weighted_token_throughput": 843.2419803433036,
  "totals": {
    "requests": 8,
    "success": 8,
    "prompt_tokens": 225499,
    "input_tokens_uncached": 58251,
    "cached_tokens": 167248,
    "completion_tokens": 1913,
    "reasoning_tokens": 1807
  },
  "targets": {
    "output_tps_p10_per_request_gte_20": false,
    "ttft_p90_lte_5": false,
    "cache_hit_rate_gte_50pct": true,
    "success_rate_gte_99pct": true,
    "weighted_token_throughput_gte_8000": false
  },
  "results": [
    {
      "ok": true,
      "elapsed_sec": 37.44546605460346,
      "ttft_sec": 8.050519423559308,
      "prompt_tokens": 7498,
      "cached_tokens": 0,
      "completion_tokens": 256,
      "reasoning_tokens": 196,
      "output_tps": 8.708979921386797,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 42.67267605289817,
      "ttft_sec": 23.193294195458293,
      "prompt_tokens": 34177,
      "cached_tokens": 7488,
      "completion_tokens": 121,
      "reasoning_tokens": 75,
      "output_tps": 6.21169608386653,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 44.97984442859888,
      "ttft_sec": 3.3009007796645164,
      "prompt_tokens": 35384,
      "cached_tokens": 34160,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 6.142190218550449,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 39.61451355740428,
      "ttft_sec": 9.986752737313509,
      "prompt_tokens": 19274,
      "cached_tokens": 7472,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 8.640544979234637,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 47.399366622790694,
      "ttft_sec": 3.8023243956267834,
      "prompt_tokens": 37099,
      "cached_tokens": 35376,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 5.87195797976622,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 36.7238706368953,
      "ttft_sec": 6.21262020803988,
      "prompt_tokens": 14661,
      "cached_tokens": 7472,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 8.390347704592697,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 47.17942987754941,
      "ttft_sec": 3.504132028669119,
      "prompt_tokens": 38199,
      "cached_tokens": 37088,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 5.86143684436403,
      "error": null
    },
    {
      "ok": true,
      "elapsed_sec": 46.50449903123081,
      "ttft_sec": 3.3495454359799623,
      "prompt_tokens": 39207,
      "cached_tokens": 38192,
      "completion_tokens": 256,
      "reasoning_tokens": 256,
      "output_tps": 5.932111581002198,
      "error": null
    }
  ]
 }
--- a/worklogs/remote_scripts/start_platform_245k_repro.sh
+++ b/worklogs/remote_scripts/start_platform_245k_repro.sh
@@ -4,6 +4,8 @@ set -euo pipefail
 OUT="${1:-/root/work/logs/platform_245k_repro_20260715_01}"
 PORT="${PORT:-1111}"
 MODEL="${MODEL:-/root/public-storage/models/Qwen/Qwen3.6-35B-A3B}"
 MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-16384}"
 MOE_DECODE_NATIVE="${MOE_DECODE_NATIVE:-0}"
 mkdir -p "$OUT"
 cd /root
@@ -18,7 +20,7 @@ export RMSNORM_NATIVE=1
 export LOOP1_NATIVE=1
 export PROF_TRACE=0
 export PROF_MOE_SUMMARY=0
-export MOE_DECODE_NATIVE=0
+export MOE_DECODE_NATIVE
 python3 - <<'PY'
 import os
@@ -55,6 +57,7 @@ print("sigkill_pids", left)
 PY
 echo "[repro] gpu before start" | tee "$OUT/driver.log"
 echo "[repro] MAX_NUM_BATCHED_TOKENS=$MAX_NUM_BATCHED_TOKENS MOE_DECODE_NATIVE=$MOE_DECODE_NATIVE" | tee -a "$OUT/driver.log"
 LD_LIBRARY_PATH=/usr/local/corex-3.2.3/lib64:/usr/local/corex/lib64:/usr/local/corex/lib:/usr/local/iluvatar/lib64 \
  /usr/local/corex/bin/ixsmi | tee -a "$OUT/driver.log"
@@ -71,7 +74,7 @@ nohup python3 -m vllm.entrypoints.openai.api_server \
  --max-num-seqs 1 \
  --disable-log-requests \
  --disable-frontend-multiprocessing \
-  --max-num-batched-tokens 16384 \
+  --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" \
  --enable-chunked-prefill \
  --max-seq-len-to-capture 245000 \
  --enable-auto-tool-choice \