XiaomiMiMo

/

MiMo-V2-Flash

like 749

Follow

Xiaomi MiMo 3.91k

因为设备受限问题我们没有做sm100及以上 的具体部署测试，这个问题或许该去sglang 的社区问一下？

I can load the model with --attention-backend=triton

FlashInfer does not supports attention sinks

[2025-12-19 15:24:34 TP1] Capture cuda graph begin. This can take up to several minutes. avail mem=6.50 GB
[2025-12-19 15:24:34 TP0] Capture cuda graph begin. This can take up to several minutes. avail mem=4.16 GB
[2025-12-19 15:24:34 TP0] Capture cuda graph bs [1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64]
Capturing batches (bs=64 avail_mem=3.98 GB):   0%|          | 0/12 [00:10<?, ?it/s]
[2025-12-19 15:24:44 TP1] Scheduler hit an exception: Traceback (most recent call last):
  File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 2795, in run_scheduler_process
    scheduler = Scheduler(
                ^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 324, in __init__
    self.tp_worker = TpModelWorker(
                     ^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/managers/tp_worker.py", line 252, in __init__
    self._model_runner = ModelRunner(
                         ^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 362, in __init__
    self.initialize(min_per_gpu_memory)
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 534, in initialize
    self.init_device_graphs()
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 2531, in init_device_graphs
    self.graph_runner = graph_runners[self.device](self)
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 358, in __init__
    self.capture()
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 513, in capture
    _capture_one_stream()
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 500, in _capture_one_stream
    ) = self.capture_one_batch_size(bs, forward, stream_idx)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 719, in capture_one_batch_size
    run_once()
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 706, in run_once
    logits_output_or_pp_proxy_tensors = forward(
                                        ^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/models/mimo_v2_flash.py", line 772, in forward
    hidden_states, hidden_states_before_norm = self.model(
                                               ^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/models/mimo_v2_flash.py", line 648, in forward
    hidden_states, residual = layer(
                              ^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/models/mimo_v2_flash.py", line 545, in forward
    hidden_states = self.self_attn(
                    ^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/models/mimo_v2_flash.py", line 423, in forward
    attn_output = self.attn(q, k, v, forward_batch, sinks=self.attention_sink_bias)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/layers/radix_attention.py", line 123, in forward
    return forward_batch.attn_backend.forward(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/layers/attention/base_attn_backend.py", line 93, in forward
    return self.forward_decode(
           ^^^^^^^^^^^^^^^^^^^^
TypeError: FlashInferAttnBackend.forward_decode() got an unexpected keyword argument 'sinks'

建议使用fa3 backend: --attention-backend=fa3

FA3 does not supports RTX Pro 6000

不支持哈，rtx pro 6000是blackwell架构，FA3只支持hopper