GpuPassthroughRemediationProvider returned early unless the NVIDIA
runtime was registered with Docker. AMD doesn't register one, so the
log-based CPU-fallback detection added in v1.35.0-rc.1 never ran on
AMD boxes. On a Phoenix1 iGPU running ollama:rocm with /dev/kfd and
/dev/dri passed through, Ollama logged library=cpu while gpuHealth
reported status "ok".
The gpuHealth side had its own bug. hasNvidiaRuntime and
hasRocmRuntime were set only inside the lspci probe branch. When lspci
returned a plausible controller, the else branch reported "ok" without
probing and left both flags false. A blind reinstall would not have
helped either: gfx1103 needs HSA_OVERRIDE_GFX_VERSION=11.0.0, and
without the gfx marker or ai.amdHsaOverride a reinstall rebuilds the
same CPU-bound container.
- Add resolveExpectedGpuVendor(). It reads the nomad_ollama container
config first (nvidia DeviceRequest, ROCm image tag, /dev/kfd), then
the nvidia runtime, then the AMD GPU type when no container exists.
/dev/dri alone is not treated as AMD. A CPU container on an AMD box
is trusted as deliberate (#1232 fallback, acceleration disabled).
- Gate the provider on the expected vendor. NVIDIA auto-reinstall is
unchanged but now skips when the container requests a GPU with no
runtime registered. AMD is detect-only: the provider logs whether a
reinstall would change the container and never reinstalls itself.
- In gpuHealth, set the runtime flags up front and let Ollama's
startup log decide health whenever a GPU is expected, whatever lspci
returned. A CPU-only log now reports passthrough_failed and skips
the nvidia-smi exec. lspci stays the display source when clean.
- Parse the first GPU "inference compute" line instead of the last
line, which could be the CPU fallback device, and recognize Vulkan.
Move the startup-log reader into ollama_compute so both callers
share it.
- Report amdReinstallWouldChange, amdGfxTarget, currentHsaOverride and
suggestedHsaOverride on AMD passthrough_failed. The gfx target is
parsed from Ollama's "amdgpu is not supported" line (best-effort).
- Add GpuPassthroughAlert, shared by Settings > System and Settings >
Models. When a reinstall would rebuild the same AMD container, it
offers "Fix: Set GFX Override" instead of a reinstall. The dialog
saves ai.amdHsaOverride and then reinstalls.
- Expose ai.amdHsaOverride through the settings API. The value lands
in the container env, so only x.y.z, "none" or empty is accepted.
- Silence the HSA override resolver's logging on the gpuHealth poll.
The #1325 case where no GPU runtime is expected and lspci reports a
controller still returns "ok" unprobed.
Fixes#1344
Refs #1325, #1232
Chat sent num_ctx, num_predict and nothing else, so every turn decoded at
whatever the backend picked: temperature 0.8, top_p 0.9, no min_p. On a
heavily quantized 1-3B model that is the wandering, repetitive output people
read as "the local model is bad".
Adds `ai.responseStyle` with four profiles. min_p is the lever, not
temperature: it floors relative to the top token, so it truncates hard on a
confident distribution and barely at all on a flat one, which is where small
models start to wander. top_p stays at 1.0 rather than stacked on top of it.
repeat_penalty moves with the style, because a grounded answer legitimately
repeats a part number inside the 64-token window: `focused` relaxes it to
1.05, `creative` tightens it to 1.15.
`auto` is the only style that defers to the model. Plenty of library models
ship author-tuned samplers (qwen3 sets 0.6/0.95/20), so those win key by key
and NOMAD fills only the gap, which in practice is min_p. A style the user
picked by name overrides them, or it would silently do nothing on half
their models.
The compat transport drops min_p, so a profile carries a separate compatTopP.
Passing top_p 1.0 to a backend with no min_p behind it would sample the raw
distribution, which is worse than the 0.9 it replaced.
Resolved in the controller, so the grammar-constrained task calls keep their
temperature 0. An explicit temperature still outranks a profile's, which is
what keeps the eval harness deterministic.
`eval:generation --style` is new and load-bearing: temperature 0 makes min_p,
top_p and top_k inert, so a default run cannot see this change at all.
`--style=off` sends no samplers and no temperature, and is the control.
Verified against Ollama 0.32.9, min_p accepted. 22 unit tests.
There was one cutoff, Qdrant's score_threshold at 0.3, and nomic scores
unrelated text well above it, so every question retrieved *something*
(nonEmptyRateOnRefusal 1.0) and rag_context wasted four lines asking a 3B
model to "silently judge" relevance. That is a classification a number
should decide, not a micro model.
Adds `RAG_MIN_FINAL_SCORE`, a floor on the post-rerank score, applied before
source diversity. The floor judges relevance, the 0.85^n penalty judges
redundancy. Nothing clears it and nothing is injected. `score_threshold` stays
at 0.3 as the candidate net. Exposed as `rag.minRelevance` with a module-
scoped TTL cache invalidated from `updateSetting`.
The floor is calibrated and the harness was describing the wrong number:
aggregate() bucketed semanticScore, so its percentiles measured raw cosine
while the new knob filters the reranked score. Both axes now, and the gap
matters: recall falls at 0.60 on the semantic axis, 0.66 on the final one.
floor recall@5 precision@5 emptyOnAnswerable nonEmptyOnRefusal
0.00 0.991 0.221 0% 100%
0.62 0.991 0.524 0% 60%
0.66 0.991 0.695 0% 60%
0.67 0.972 0.704 1% 60%
Default 0.62, not the last-lossless 0.66: the refusal win is complete at
0.62 and 28 documents is a thin basis for spending margin. 0.60 is the
honest floor. Three of five refusal goldens are near-corpus by
construction, where declining is the generation tier's job.
Both eval tiers pin the floor to the constant and ignore the setting, for
the same reason the harness never sets skipRetrieval.
4 metrics improved against baseline, no regressions.
Chat was run through Ollama's OpenAI-compatible /v1 endpoint, which has no
context-size field, so the num_ctx we sent was silently discarded and the
entire NUM_CTX_LADDER was dead code. Every conversation ran at the backend
default and was truncated from the middle (where history and
retrieved context actually sit). Measured with identical ~7k-token prompts:
2,060 tokens processed against 7,049 on the native endpoint, and with the
needed fact in the discarded middle the old path hallucinated a plausible
wrong number.
Transport: add a native /api/chat path using the already-installed ollama
package, sending a real options object (num_ctx, num_predict, temperature,
seed) plus keep_alive. The OpenAI path is still used for LM Studio and llama.cpp,
minus the ineffective num_ctx and with include_usage so both paths return
prompt token counts. /api/show is now read for trained context, parameter
size, quantization and any modelfile num_ctx. The <think> stream normalizer
moves to app/utils/think_stream.ts so both transports share it; it gains a
flush() that fixes a dangling partial tag being dropped at end of stream.
Token accounting: replace chars/N with a structural segmenter scaled by a
per-model factor learned by EWMA from the prompt_eval_count every response
already returned and threw away. End-to-end estimator error is 4.6% against
18-20% for a fixed divisor — qwen2.5 converges to k=1.26 where llama3 sits
at 1.00, so no single global constant serves both.
Context window: resolved once per model from trained context, the exact KV
cost derived from GGUF metadata (head_count_kv, not head_count — GQA makes
them differ 4x), and available VRAM, then memoized. Memoization is
load-bearing: asking for a different num_ctx makes Ollama unload and reload
the model, stalling the turn and discarding the KV cache. The floor never
overrides the model's trained ceiling — tinyllama trains to 2048, and a
blind 4096 floor pushes RoPE past anything the weights saw.
Prompt budget: planPrompt allocates explicitly instead of delegating
overflow to silent backend truncation — response reserve, then fixed blocks
(system and current query, never dropped), then RAG chunks whole and
best-first, then history in whole user+assistant turns, evicted newest-first
in blocks with an elision marker. Block eviction keeps the prefix stable for
several turns rather than shifting on every one.
KV-cache ordering: the per-turn retrieved block moves behind history to the
tail, so the stable prefix stays byte-identical turn to turn. At turn 14 of
a long conversation, prefill is 231ms at the tail against 539ms at the front
(~1ms/turn growth against ~25ms). The win is conditional — with a large
retrieved block and short history the two measure the same — so both
orderings ship behind RAG_PLACEMENT. The per-turn query rewrite also moves
to the tasks model; on a single-slot Ollama it was evicting the chat model's
cached prefix on every single turn.
Also: OLLAMA_CONTEXT_LENGTH and OLLAMA_KV_CACHE_TYPE=q8_0 on the container
(q8_0 is treated as headroom, never as a budgeting assumption, since
unsupported architectures fall back to f16 silently), a Context Window
setting in Settings > Models, and budget fields on PipelineTrace.
Eval: golden suites (--suite) so the new long_context goldens cannot move
the aggregates the committed baselines were recorded against, plus budget
and estimator-error metrics through the report path. Retrieval is unchanged
and eval:compare reports no metric outside the tolerance band; the
long_context suite scores correctness 1.000 with historyElided 0.500 (the
two eviction cases, by design). Unit tests 337 -> 386 passing, same 7
pre-existing failures.
Ingest-side CHAR_TO_TOKEN_RATIO is deliberately untouched: changing it
re-chunks the corpus and invalidates retrieval baselines, which belongs in
its own measured change.
Adds a self-update path for the NOMAD admin/core image that runs without a
human in the loop, gated by layered safety checks. Recreation remains the
sidecar's job; this only decides *whether* to update now and requests it.
An hourly job (AutoUpdateJob) evaluates a side-effect-free decision pipeline
(AutoUpdateService.evaluate):
- opt-in: disabled by default
- in a user-configured local-time window (handles midnight wrap)
- an eligible release: same major (major bumps stay manual), strictly newer,
past a configurable cool-off, GA only (no drafts/prereleases), strict semver
- pre-flight: sidecar present, no in-flight system update / downloads / app
installs, and sufficient host disk (estimated from the registry manifest)
When all pass, it drives SystemUpdateService.requestUpdate() with a vetted tag.
Failure backoff auto-disables after 3 genuine update-request failures; transient
release-lookup errors are skips (offline-first appliances are routinely without
connectivity). Re-enabling clears the backoff state.
Settings UI exposes the toggle, window, and cool-off, with live status (eligible
target, in/out of window, last result/error). A `node ace auto-update:dry-run`
command exercises the full pipeline — and a deterministic --scenarios suite the
pure decision logic — without ever triggering an update.
New KVStore keys under autoUpdate.* hold config + last-run state.
