Summary

Two one-line fixes that make the server actually run on x86_64 Linux. Before these patches the server aborts on every run — the model was completely unusable on Linux.

Tested on: Ubuntu 24.04, x86_64, GCC 13.3, CUDA 12.1, RTX 3090, DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2.gguf.

Bug 1 — sched_reserve assert when n_ctx <= n_batch

File: src/llama-context.cpp

sched_reserve() picks a synthetic start position reserve_pos0 to size the "resumed decode" graph. When n_ctx <= n_tokens the fallback was n_tokens, placing the dry-run batch at positions [n_tokens .. 2*n_tokens-1] — entirely outside the KV cache [0 .. n_ctx-1].

Inside llm_build_deepseek4 the decode path then computed n_comp_visible = 2*n_ctx/ratio vs n_comp_cache = n_ctx/ratio, triggering:

GGML_ASSERT(n_comp_visible <= n_comp_cache) failed

common_params_fit_impl() probes n_ctx = n_batch as its first candidate, so this assertion fires unconditionally before inference ever starts.

Fix: use 0 as the fallback. When n_ctx == n_batch a "resumed" full-batch decode is impossible anyway — prefill-from-zero is the only valid graph shape.

Bug 2 — CUDA GGML_OP_CONCAT dispatch mismatch

File: ggml/src/ggml-cuda/ggml-cuda.cu

ggml_backend_cuda_device_supports_op() returned true for GGML_OP_CONCAT with any type except I32/I16. But ggml_cuda_op_concat() asserts src0->type == GGML_TYPE_F32. DeepSeek V4 attention state tensors are non-F32, so the backend scheduler dispatched them to CUDA and the kernel aborted on the first decode step:

concat.cu:165: GGML_ASSERT(src0->type == GGML_TYPE_F32) failed

Fix: restrict supports_op for GGML_OP_CONCAT to F32 only, matching what the kernel actually implements.