Skip to content

R9700/AMD Ryzen AI Max 395 iGPU (Radeon 8060S) device number misidentified #712

Description

@griffin-warrior

podman run --rm --device /dev/kfd --device /dev/dri --group-add video --group-add render --security-opt seccomp=unconfined -p 8080:8080 -v "$PWD/models:/opt/lucebox-hub/server/models" ghcr.io/luce-org/lucebox-hub:rocm

[INFO] Wrote /opt/lucebox-hub/HOST_INFO (source=unknown)
[INFO] Auto-detected target: DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf
[INFO] Resolved draft dir /opt/lucebox-hub/server/models/draft → /opt/lucebox-hub/server/models/draft/DeepSeek-V4-Flash-0731-DSpark-draft-Q4RMFP4-denseF16.gguf
[INFO] lucebox-hub container starting (target=DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf, max_ctx=16384, budget=22, lazy=0)
[server] warning: --draft ignored: architecture 'deepseek4' has no speculative decode support
[server] warning: --ddtree ignored: architecture 'deepseek4' has no DDTree speculative decode support
[server] loading tokenizer from /opt/lucebox-hub/server/models/DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf
[tokenizer] added_tokens: 1283 special tokens
[tokenizer] loaded vocab=129280 merges=127741 bos=0 eos=1 eot=-1 pre=qwen2 sp=no
[server] creating backend...
[backend_factory] detected arch=deepseek4
[deepseek4] gfx1201 hybrid prefill: defaulting hot expert sub-batch to 4
ggml_cuda_init: found 2 ROCm devices (Total VRAM: 130928 MiB):
Device 0: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32, VRAM: 32624 MiB
Device 1: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 98304 MiB
[deepseek4] heterogeneous target detected; using hybrid expert load path
[deepseek4] model: layers=43 embd=4096 heads=64 head_dim=512 lora_q=1024 lora_o=1024 out_groups=8
[deepseek4] moe: experts=256 used=6 shared=1 ff=2048 hash_layers=3
[deepseek4] attention: swa=128 rot=64 indexer_heads=64 top_k=512 hc=4
[deepseek4] qtype-106 experts not resident on this shard (skip_expert_tensors) — fused decode disabled here
[deepseek4] qtype-105 down-experts not resident on this shard (skip_expert_tensors) — fused decode disabled here
[deepseek4] loaded 1328 tensors, 3593.2 MB GPU buffer, 0.0 MB dense TP split
[deepseek4] hybrid placement: gpu_total=31.86 GiB gpu_free=28.14 GiB core=3.72 GiB kv=0.13 GiB warm=0.25 GiB safety=0.50 GiB expert_budget=27.26 GiB hot/layer=79
[deepseek4] placement expert_memory: total=87.62 GiB uniform_round=350.50 MiB hot=3397 27.04 GiB cold=7611 60.58 GiB
[deepseek4] qtype-106 (mixed ROCmFP2) gate experts cannot decode from hybrid/cold placement; falling back to monolithic full load
[deepseek4] model: layers=43 embd=4096 heads=64 head_dim=512 lora_q=1024 lora_o=1024 out_groups=8
[deepseek4] moe: experts=256 used=6 shared=1 ff=2048 hash_layers=3
[deepseek4] attention: swa=128 rot=64 indexer_heads=64 top_k=512 hc=4
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 93321.23 MiB on device 0: cudaMalloc failed: out of memory
[deepseek4] monolithic fallback failed (model does not fit resident): /opt/lucebox-hub/server/models/DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf
[deepseek4] hybrid mode failed: /opt/lucebox-hub/server/models/DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf
[backend_factory] DeepSeek4Backend init failed
[server] backend creation failed

It looks like dflash_server found the wrong device.

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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions