Skip to content

server: LFM2.5 MoE with DSpark aborts during CPU context init #41

Description

@am17an

Upstream: ggml-org#28614

LFM2.5-8B-A1B with its DSpark draft aborts in graph reservation before the server starts when both models are on CPU.

Environment: Linux aarch64 (GB10 host), current main f3f1a8f2760f28325a5ec20c05b171e5b7c83a29; exact LiquidAI Q4_K_M target and Q8_0 DSpark GGUFs.

llama-server \
  -m LFM2.5-8B-A1B-Q4_K_M.gguf \
  -md LFM2.5-8B-A1B-DSpark-Q8_0.gguf \
  -ngl 0 -ngld 0 \
  --spec-type draft-dspark --spec-draft-n-max 10 --spec-draft-n-min 0 \
  --ctx-size 4096 --flash-attn on

Observed:

ggml-backend.cpp:1956: GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs) failed

The same command without -ngl 0 -ngld 0 loads successfully on CUDA. The immediate cause is an undersized scheduler hash set while reserving the CPU graph for this MoE + DSpark configuration; introducing commit not identified.

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

    bugSomething isn't workingserver

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions