Skip to content

Cant load qwen 3.8 27B in gpu 5090 24gb mobile #417

Description

@rioryugen

Before you start

  • I have read the FAQ and my problem is not answered there.
  • I have read the Roadmap and this is not already planned there.
  • I have searched existing issues and found no duplicate.
  • I have restarted the Desktop app to pick up the latest update and the problem still happens.

What happened

[22:35:47] cmd/INFO ft serve --model D:\ai_model\Qwen3.8-27B-NVFP4 --port 1919 --max-running-requests 4 --memory-ratio 0.85 --host 127.0.0.1 --cors-origins tauri://localhost,http://tauri.localhost,http://localhost:1420 --gpu GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9
[22:35:47] stdout/INFO serve started (pid=34784 model=D:\ai_model\Qwen3.8-27B-NVFP4 port=1919)
[22:35:51] stdout/INFO [2026-09-08|22:35:51] INFO Parsed arguments:
[22:35:51] stdout/INFO ServerArgs(model_path='D:\ai_model\Qwen3.8-27B-NVFP4', tp_info=DistributedInfo(rank=0, size=1), dtype=torch.bfloat16, max_running_req=4, attention_backend='auto', moe_backend='auto', nvfp4_backend='triton', ple_backend='disk', expert_load='auto', moe_cache_size=0, moe_cache_rate=None, moe_cache_auto=False, kv_reserve_tokens=8192, moe_cache_policy='lru', moe_prefill_overlap=True, moe_prefill_hit_d2d=False, moe_collect_stats=False, moe_cpu_threads=0, moe_cpu_layers=None, moe_hybrid_max_fetch=-1, cuda_graph_bs=None, cuda_graph_max_bs=None, page_size=1, memory_ratio=0.85, linear_state_cache_ratio=2.0, swa_full_tokens_ratio=0.2, swa_num_pages_override=None, distributed_timeout=60.0, use_dummy_weight=False, use_pynccl=True, max_seq_len_override=None, num_page_override=None, num_token_override=None, max_extend_tokens=8192, cache_type='radix', offline_mode=False, decode_log_interval=40, special_token_ckpt=False, _unique_suffix='.pid=22772', _ipc_base_port=51266, server_host='127.0.0.1', server_port=1919, num_tokenizer=0, silent_output=False, shell_mode=False, served_model_name='Qwen3.8-27B-NVFP4', tool_call_parser='qwen3_coder', reasoning_parser='qwen3', sampling_defaults='model', max_output_tokens=None, enable_cache_report=False, cors_origins='tauri://localhost,http://tauri.localhost,http://localhost:1420', gpu=('GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9',), gpu_assigned=None)
[22:35:51] stdout/INFO [2026-09-08|22:35:51|initializer] INFO --gpu GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9 -> GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9
[22:35:51] stdout/INFO [2026-09-08|22:35:51|FrontendAPI] INFO Default sampling config (source=model): temperature=1.0, top_k=20, top_p=0.95
[22:35:51] stdout/INFO INFO: Started server process [22772]
[22:35:51] stdout/INFO INFO: Waiting for application startup.
[22:35:51] stdout/INFO INFO: Application startup complete.
[22:35:51] stdout/INFO INFO: Uvicorn running on http://127.0.0.1:1919 (Press CTRL+C to quit)
[22:35:54] stdout/INFO C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py:81: FutureWarning: torch.cuda._set_allocator_settings is deprecated. Use torch._C._accelerator_setAllocatorSettings instead.
[22:35:54] stdout/INFO scheduler = Scheduler(args)
[22:35:54] stdout/INFO [2026-09-08|22:35:54|core|rank=0] INFO Enabled expandable_segments (override via PYTORCH_ALLOC_CONF)
[22:35:54] stdout/INFO [2026-09-08|22:35:54|core|rank=0] INFO Auto-selected attention backend: triton
[22:35:54] stdout/INFO [2026-09-08|22:35:54|core|rank=0] INFO Resolved config: attention_backend='triton', cache_type='hybrid_radix', page_size=1
[22:35:54] stdout/INFO [2026-09-08|22:35:54|core|rank=0] INFO Free memory before loading model: 22.60 GiB
[22:35:54] stdout/INFO C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\torch\utils_device.py:116: UserWarning: expandable_segments not supported on this platform (Triggered internally at C:\actions-runner_work\pytorch\pytorch\pytorch\c10/cuda/CUDAAllocatorConfig.h:39.)
[22:35:54] stdout/INFO return func(*args, **kwargs)
[22:36:06] stdout/INFO Loading weights (FTW): 98%|█████████▊| 18.3G/18.8G [00:11<00:00, 1.85GB/s]
[22:36:07] stdout/INFO
[22:36:07] stdout/INFO Process freetoken-TP0-scheduler:
[22:36:07] stdout/ERROR [2026-09-08|22:36:07|FrontendAPI] ERROR Backend supervisor: AssertionError: Not enough memory for KV cache, try reducing --num-pages
[22:36:07] stdout/INFO Traceback (most recent call last):
[22:36:07] stdout/INFO File "C:\Users\mnkna\AppData\Roaming\uv\python\cpython-3.12-windows-x86_64-none\Lib\multiprocessing\process.py", line 314, in _bootstrap
[22:36:07] stdout/INFO self.run()
[22:36:07] stdout/INFO File "C:\Users\mnkna\AppData\Roaming\uv\python\cpython-3.12-windows-x86_64-none\Lib\multiprocessing\process.py", line 108, in run
[22:36:07] stdout/INFO self._target(*self._args, **self._kwargs)
[22:36:07] stdout/INFO File "C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py", line 81, in _run_scheduler
[22:36:07] stdout/INFO scheduler = Scheduler(args)
[22:36:07] stdout/INFO ^^^^^^^^^^^^^^^
[22:36:07] stdout/INFO File "python/freetoken/scheduler/scheduler.py", line 65, in freetoken.scheduler.scheduler.Scheduler.init
[22:36:07] stdout/INFO File "python/freetoken/engine/engine.py", line 391, in freetoken.engine.engine.Engine.init
[22:36:07] stdout/INFO File "C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\kvcache\base.py", line 73, in solve_num_pages
[22:36:07] stdout/INFO assert num_pages > 1, "Not enough memory for KV cache, try reducing --num-pages"
[22:36:07] stdout/INFO ^^^^^^^^^^^^^
[22:36:07] stdout/INFO AssertionError: Not enough memory for KV cache, try reducing --num-pages
[22:36:15] health/ERROR AssertionError: Not enough memory for KV cache, try reducing --num-pages (×5)
[22:36:16] stdout/INFO INFO: 127.0.0.1:60374 - "POST /v1/admin/prepare-stop HTTP/1.1" 200 OK
[22:36:16] stdout/INFO serve stopped (pid=34784)
[22:36:55] cmd/INFO ft serve --model D:\ai_model\Qwen3.8-27B-NVFP4 --port 1919 --max-running-requests 4 --memory-ratio 0.93 --host 127.0.0.1 --cors-origins tauri://localhost,http://tauri.localhost,http://localhost:1420 --gpu GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9
[22:36:55] stdout/INFO serve started (pid=44496 model=D:\ai_model\Qwen3.8-27B-NVFP4 port=1919)
[22:37:00] stdout/INFO [2026-09-08|22:37:00] INFO Parsed arguments:
[22:37:00] stdout/INFO ServerArgs(model_path='D:\ai_model\Qwen3.8-27B-NVFP4', tp_info=DistributedInfo(rank=0, size=1), dtype=torch.bfloat16, max_running_req=4, attention_backend='auto', moe_backend='auto', nvfp4_backend='triton', ple_backend='disk', expert_load='auto', moe_cache_size=0, moe_cache_rate=None, moe_cache_auto=False, kv_reserve_tokens=8192, moe_cache_policy='lru', moe_prefill_overlap=True, moe_prefill_hit_d2d=False, moe_collect_stats=False, moe_cpu_threads=0, moe_cpu_layers=None, moe_hybrid_max_fetch=-1, cuda_graph_bs=None, cuda_graph_max_bs=None, page_size=1, memory_ratio=0.93, linear_state_cache_ratio=2.0, swa_full_tokens_ratio=0.2, swa_num_pages_override=None, distributed_timeout=60.0, use_dummy_weight=False, use_pynccl=True, max_seq_len_override=None, num_page_override=None, num_token_override=None, max_extend_tokens=8192, cache_type='radix', offline_mode=False, decode_log_interval=40, special_token_ckpt=False, _unique_suffix='.pid=25976', _ipc_base_port=51322, server_host='127.0.0.1', server_port=1919, num_tokenizer=0, silent_output=False, shell_mode=False, served_model_name='Qwen3.8-27B-NVFP4', tool_call_parser='qwen3_coder', reasoning_parser='qwen3', sampling_defaults='model', max_output_tokens=None, enable_cache_report=False, cors_origins='tauri://localhost,http://tauri.localhost,http://localhost:1420', gpu=('GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9',), gpu_assigned=None)
[22:37:00] stdout/INFO [2026-09-08|22:37:00|initializer] INFO --gpu GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9 -> GPU-4efc32ff-f2d9-2d01-aa01-936583b77aa9
[22:37:00] stdout/INFO [2026-09-08|22:37:00|FrontendAPI] INFO Default sampling config (source=model): temperature=1.0, top_k=20, top_p=0.95
[22:37:00] stdout/INFO INFO: Started server process [25976]
[22:37:00] stdout/INFO INFO: Waiting for application startup.
[22:37:00] stdout/INFO INFO: Application startup complete.
[22:37:00] stdout/INFO INFO: Uvicorn running on http://127.0.0.1:1919 (Press CTRL+C to quit)
[22:37:03] stdout/INFO C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py:81: FutureWarning: torch.cuda._set_allocator_settings is deprecated. Use torch._C._accelerator_setAllocatorSettings instead.
[22:37:03] stdout/INFO scheduler = Scheduler(args)
[22:37:03] stdout/INFO [2026-09-08|22:37:03|core|rank=0] INFO Enabled expandable_segments (override via PYTORCH_ALLOC_CONF)
[22:37:03] stdout/INFO [2026-09-08|22:37:03|core|rank=0] INFO Auto-selected attention backend: triton
[22:37:03] stdout/INFO [2026-09-08|22:37:03|core|rank=0] INFO Resolved config: attention_backend='triton', cache_type='hybrid_radix', page_size=1
[22:37:03] stdout/INFO [2026-09-08|22:37:03|core|rank=0] INFO Free memory before loading model: 22.60 GiB
[22:37:03] stdout/INFO C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\torch\utils_device.py:116: UserWarning: expandable_segments not supported on this platform (Triggered internally at C:\actions-runner_work\pytorch\pytorch\pytorch\c10/cuda/CUDAAllocatorConfig.h:39.)
[22:37:03] stdout/INFO return func(*args, **kwargs)
[22:37:15] stdout/INFO Loading weights (FTW): 99%|█████████▉| 18.7G/18.8G [00:11<00:00, 1.52GB/s]
[22:37:15] stdout/INFO Process freetoken-TP0-scheduler:
[22:37:15] stdout/ERROR [2026-09-08|22:37:15|FrontendAPI] ERROR Backend supervisor: AssertionError: Not enough memory for KV cache, try reducing --num-pages
[22:37:15] stdout/INFO Traceback (most recent call last):
[22:37:15] stdout/INFO File "C:\Users\mnkna\AppData\Roaming\uv\python\cpython-3.12-windows-x86_64-none\Lib\multiprocessing\process.py", line 314, in _bootstrap
[22:37:15] stdout/INFO self.run()
[22:37:15] stdout/INFO File "C:\Users\mnkna\AppData\Roaming\uv\python\cpython-3.12-windows-x86_64-none\Lib\multiprocessing\process.py", line 108, in run
[22:37:15] stdout/INFO self._target(*self._args, **self._kwargs)
[22:37:15] stdout/INFO File "C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\server\launch.py", line 81, in _run_scheduler
[22:37:15] stdout/INFO scheduler = Scheduler(args)
[22:37:15] stdout/INFO ^^^^^^^^^^^^^^^
[22:37:15] stdout/INFO File "python/freetoken/scheduler/scheduler.py", line 65, in freetoken.scheduler.scheduler.Scheduler.init
[22:37:15] stdout/INFO File "python/freetoken/engine/engine.py", line 391, in freetoken.engine.engine.Engine.init
[22:37:15] stdout/INFO File "C:\Users\mnkna\AppData\Local\FreeToken\venv\Lib\site-packages\freetoken\kvcache\base.py", line 73, in solve_num_pages
[22:37:15] stdout/INFO assert num_pages > 1, "Not enough memory for KV cache, try reducing --num-pages"
[22:37:15] stdout/INFO ^^^^^^^^^^^^^
[22:37:15] stdout/INFO AssertionError: Not enough memory for KV cache, try reducing --num-pages
[22:37:17] health/ERROR AssertionError: Not enough memory for KV cache, try reducing --num-pages (×2)

Desktop app version

0.20

OS

Windows 11

OS details

Windows 1`

GPU and driver

RTX 5090 Laptop

CPU and system RAM

ultra 9 275 hx; 64gb

Checkpoint

https://huggingface.co/RadixArk/Qwen3.8-27B-NVFP4

Model settings

No response

Engine log


Anything else

No response

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Desktopproblem related to FreeToken DesktopbugSomething isn't workingwindows

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions