Reconfigure to 3 parallel 160000-ctx slots (was 1 slot). KV cache is
now asymmetric: K at q8_0, V at q5_0, both with Hadamard. The 3 slots
triple the KV, so expert blocks 27-34 are pinned back to CPU (freed
~8.7 GiB) leaving output-adjacent 35-40 on CUDA. Q8_0 model,
split-mode none, spec disabled (ik limits spec to 1 slot). AGENTS.md:
drop the scratch-port rule (test/bench processes cannot share the
single GPU with the running service anyway). Env, serve script, bats,
docs updated.
Switch to CRACK Q8_0 model (37.8 GB, mmproj f16). Servt as a single
160000-ctx slot (PARALLEL=1); speculative decoding stays disabled. KV
cache is q6_0 + Hadamard. iq5_k is not a valid KV-cache type in ik_llama
(cache whitelist excludes iq5_*); q8_KV reproduces the ggml.c:5427 view
assert even at PARALLEL=1/no-spec, so q6_0 is the verified working 6-bit
choice. Env, serve script, tests, and KB updated. Deployment was left
down after the isolation test; this commit redeploys it with the
verified Q8_0/160k/q6_0 config.
Fused MoE (fused_moe / fused_up_gate / fused_mmad) is on by default in
ik_llama - no flag needed; only -no-fmoe disables it. Slide two more
expert blocks (28,29) from CPU onto the RTX 3090 (~1.26 GiB) for the
fused-MoE run, leaving ~1.6 GiB headroom. Update override regex, test,
and KB (including a note that the headroom is now tighter than the
2-3 GiB ideal).
--jinja: ik_llama only renders a custom --chat-template-file as real Jinja
when --jinja is passed (inputs.use_jinja ? apply_jinja : non-jinja
fallback). Without it the Qwen-Sharp template's thinking/image/tool
formatting is silently dropped. Add USE_JINJA=1 -> --jinja.
Experts: the RTX 3090 had ~5.4 GiB headroom; pull expert blocks 30,32,33,34
back to CUDA (each block ~630 MiB) from CPU for lower latency. Update the
OVERRIDE_EXPERT regex, its test, and the KB with the headroom-check note.
ik_llama limits speculative decoding to a single server slot (-np 1) and
hard-exits at load if a spec stage is configured with --parallel > 1:
"Speculative decoding is currently limited to a single server slot
(-np 1)." This service keeps PARALLEL=3 (one slot per user), so set
SPEC_TYPE=none to disable speculation. It is a low-parameter MoE (3B
active of 35B), so throughput is fine without it. MTP stays disabled.
KB and tests updated.
The ik_llama fast q8_KV cache type crashes at model init on this
QWEN3MOE/MTP + multimodal arch at commit b166e269 with a ggml_view size
assert (ggml.c:5427) right after the output buffer is created, looping
the service. Empirically isolated by toggling cache type: q8_0 and f16
load and serve (smoke-tested correct output), q8_KV asserts. Switch to
q8_0 + Hadamard (K/V), which still gives the 8-bit cache + Hadamard
quality win. KB updated with the crash lesson.
--split-mode graph requires >= 2 devices in ik_llama.cpp and errors hard
at model load on a single RTX 3090 ("It is not possible to use split mode
'graph' with less than 2 devices"), looping the service. Set SPLIT_MODE
=none (the default), which works with the --override-tensor CPU expert
offload. KB updated with the two-device requirement so we do not assume
graph degrades gracefully on one GPU.
Switch frank-default-small on Dankserver from upstream llama.cpp (b10121)
to ikawrakow/ik_llama.cpp (b166e269, built with the staged pip cu13
toolkit). The model (Ornith-1.5-35B-A3B CRACK Q6_K) is unchanged and the
old binary is kept for rollback.
Runtime changes:
- KV cache quantized to q8_KV with --k-cache-hadamard/--v-cache-hadamard
- --split-mode graph (single-GPU no-op, retained per operator pref)
- migrate ngram speculation to canonical
--spec-type ngram-mod:n_min=48,n_max=64,ngram_size_n=24
(legacy --spec-ngram-mod-* are rejected in ik main)
- N_GPU_LAYERS 999 (ik -ngl rejects the literal "all"); experts stay
pinned to CPU via --override-tensor
Build: GGML_CUDA_FA_ALL_QUANTS=ON + GGML_IQK_FA_ALL_QUANTS=ON mandatory
for the quantized KV cache; recipe + reasoning baked into
docs/LLM/ikllama-ornith.md. Tests and both indices updated.
Frank