* integration: improve ability to test individual models
Add OLLAMA_TEST_MODEL env var to run integration tests against a
single model.
Enhance vision tests: multi-turn chat with cached image tokens, object
counting, spatial reasoning, detail recognition, scene understanding, OCR, and
multi-image comparison.
Add tool calling stress tests with complex agent-style prompts, large
system messages, and multi-turn tool response handling.
* review comments
Previously, a partial match within a node's edge would truncate the path
to the parent snapshot - effectively making all cache types behave as
recurrent caches. Caches with only transformer layers can rewind to
arbitrary boundary so this restores this capability to improve cache
hits
* mlx: update to HEAD on 3/23
Also fixes a few misc vendoring bugs uncovered with this first update.
This also renames the version files to make them clearer.
* CUDA Fast Gated Delta kernel
* mlx: detect eval errors and panic
On model errors or missing kernels, don't mask the error, bubble it up.
Receiving from a buffered chan error consumes the value, so only the
first caller (WaitUntilRunning, HasExited, or Close) sees the signal.
Subsequent receivers block or take the wrong branch. Replace with a
closed chan struct{} which can be received from any number of times,
and store the error in a separate field.
The stderr reader used bufio.Scanner which has a 64KB max line size.
If the subprocess wrote a line exceeding this limit, the scanner would
stop reading, the OS pipe buffer would fill, and the subprocess would
deadlock.
Replace the scanner with a statusWriter that wraps io.Copy. The writer
forwards all stderr to os.Stderr while capturing the last short line
(≤256 bytes) for error reporting, avoiding both the deadlock and the
need to buffer arbitrarily long lines.
If `OLLAMA_DEBUG_LOG_REQUESTS` is set, then on server startup a temp
folder will be created. Upon any inference request, the body will be
logged to a file in this folder, as well as a small shell script to
"replay" the request using cURL.
This is just intended for debugging scenarios, not as something to turn
on normally.
Previous xml repair for glm was a good start, but we need to go further and repair any incorrect open or closing tags
Co-authored-by: Dongluo Chen <dongluo.chen@gmail.com>
Enable multiple conversations to reuse cached computations when they
share token prefixes (e.g. the same system prompt). A prefix trie
tracks shared regions so switching between conversations only
recomputes tokens that diverge. Inactive conversation state is paged
from active GPU memory to other memory and restored on demand, with LRU
eviction to keep memory usage bounded.
Slice used cmp.Or to resolve a zero stop value to the dimension size,
intended to support open-ended slices like a[i:]. This made Slice(0, 0)
indistinguishable from Slice(), so any slice with a zero stop would
silently include the entire dimension instead of being empty.
Replace cmp.Or with an explicit End sentinel and resolve negative
indices against the dimension size, matching Python/PyTorch semantics.
Defensively handle environments without a display server to ensure signin remains usable on headless VMs and SSH sessions.
- Skip calling xdg-open when neither DISPLAY nor WAYLAND_DISPLAY is set, preventing silent failures or unexpected browser handlers
- Render the signin URL as plain text instead of wrapping it in OSC 8 hyperlink escape sequences, which can be garbled or hidden by terminals that don't support them
Claude Code sends an x-anthropic-billing-header that changes on every
request. This is embedded in the system prompt and consequently
breaks the KV cache for every request. Given the size of the prompts
that Claude Code usees, this has significant performance impact.
The OpenClaw installer requires git in addition to npm. Update the
dependency check to detect both and provide specific install guidance
for whichever dependencies are missing.
`WebSearchAnthropicWriter` expects a single object per write. The new
transparent proxy will instead send it whatever bytes it sees. This
cloud-model + local-orchestration + cloud-search is a temporary code
path, so instead of making the web search code more robust to this, I
put an adapter in the middle that will flush line-by-line to preserve
the old behavior.
Add QuantizedEmbedding and EmbeddingLayer interface so models can
use quantized embedding weights and expose tied output projections.
This change updates gemma3, glm4_moe_lite, llama, qwen3, and qwen3_5
to use the new interface.
This change adds a tensorImportTransform interface for model-specific
tensor transformations during safetensors import. This allows importing
and modifying the standard HF based weights as well as the mlx-community
derived pre-quantized safetensors repos to be directly
imported into `ollama create`. Right now this only works with Qwen3.5
importing which does tensor renaming, norm weight shifting (it
adds +1 to each value of the norm vectors), conv1d transposition,
and casts to BF16s for F32 based vectors.
MLX runners (image generation and LLM) previously bypassed the
scheduler's standard load path via a separate loadMLX method. This meant
they skipped VRAM fitting checks and couldn't participate in model
eviction.
Now all model types flow through the same load function. Model eviction
for MLX is based on weights as KV cache and compute graph are dynamic.
This means that eviction does not take into account the worst case
memory and models can still compete for memory but it is a significant
improvement.
In allocModel(), the first call to reserveWorstCaseGraph(true) had its
error silently discarded — `return nil` was used instead of `return err`.
This meant that if the prompt-sized graph reservation failed (e.g. due
to insufficient memory), the error was swallowed, allocModel reported
success, and the model appeared to load correctly. Subsequent inference
would then fail in unexpected ways because the worst-case graph was
never properly reserved.
Fix: return the actual error so the caller can handle the failure
(retry with reduced parallelism, report OOM, etc.).
Co-Authored-By: Claude (claude-opus-4-6) <noreply@anthropic.com>
In container environments without systemd, `openclaw onboard
--install-daemon` exits non-zero because it cannot create a systemd
user service. This causes `ollama launch openclaw` to abort even
though the gateway can be started as a foreground child process.
Only pass --install-daemon when systemd user services are reachable
(Linux with /run/systemd/system present and XDG_RUNTIME_DIR set).
On all other platforms the flag is still included by default.
New features:
- Warmup phase to eliminate cold-start outliers
- time-to-first-token measured in each epoch
- VRAM/memory tracking to identify CPU spillover
- Controlled prompt length
- Defaults to 6 epochs and 200 tokens max
Benchstat fixes:
- ns/request instead of ns/op — non-standard unit created a separate group instead of grouping with timing metrics
- Token count as the N field — benchstat interprets N as iteration count for statistical weighting, not as a token count
OpenClaw now accepts the Ollama onboarding flags directly upstream, so rely on its wizard state instead of the legacy integration onboarding flag.
Update first-run setup to pass the Ollama auth and model flags during onboarding, perform a best-effort update before onboarding when needed, and drop the stale test that asserted persistence of the old onboarding flag.
When a zstd-compressed request (e.g. from Codex CLI) hits /v1/responses
with a cloud model the request failed.
Fix by decompressing zstd bodies before
model extraction, so cloud models are detected and proxied directly
without the writer being wrapped.
writeError in both OpenAI and Anthropic middleware writers would return
a raw json.SyntaxError when the error payload wasn't valid JSON (e.g.
"invalid character 'e' looking for beginning of value"). Fall back to
using the raw bytes as the error message instead.
Also use the actual HTTP status code rather than hardcoding 500, so
error types map correctly
Root cause: StreamConverter.Process() only incremented contentIndex when
closing a thinking block if text content was present. When a model emitted
thinking followed directly by a tool_use block (no text in between),
thinkingDone was never set and contentIndex was not incremented, causing the
tool_use content_block_start to reuse index 0. Clients expecting sequential
indices would then fail to find the tool content block.
Fix: In the tool call loop, close any open thinking block (thinkingStarted &&
!thinkingDone) and increment contentIndex before opening the tool_use block,
mirroring the existing logic that closes an open text block.
Fixes#14816
Add reasoning_effort and reasoning to the supported features and
request fields for /v1/chat/completions. These fields control
thinking on thinking-capable models but were previously undocumented.
Closes#14820
Use 7z compression (better compression rate) if found in path. That
alone isn't sufficient to get us under 2G, so MLX is now split out as a
discrete download. Fix CI so it will fail if artifacts fail to upload.
