The existing strict gemma4 tool parser is still the primary path, but if
this fails, we try to repair by fixing some of the most commonly seen
mistakes these models seem to make in practice.
We repair by building up a set of candidates, and use the first candidate
that parses.
Repairs cover:
- missing Gemma string delimiters
- single-quoted string values, including a dangling Gemma delimiter
- raw terminal string values (if the corresponding tool schema indicates
it should be a string)
- missing object close only after a concrete repair
Add regression coverage for malformed tool calls from issue #15315 and
focused unit tests for the individual repair helpers and candidate
pipeline.
We've observed Gemma 4 occasionally emitting extra <tool_call|> tags
after a valid tool call. We suppress leading close tags in this
immediate post-tool-call state so the extra close tags do not leak into
assistant content. The tradeoff is that if the model intentionally
begins its next content span with the literal string "<tool_call|>", we
will erroneously treat it as noise and drop it.
Replace the custom Gemma4 argument normalizer with a stricter
reference-style conversion: preserve Gemma-quoted strings, quote bare
keys, and then unmarshal the result as JSON.
This keeps quoted scalars as strings, preserves typed unquoted values,
and adds test coverage for malformed raw-quoted inputs that the
reference implementation rejects.
cublasGemmBatchedEx fails during graph capture when pool allocations
return fake pointers. This is triggered when NUM_PARALLEL is greater
than 1 for models like gemma4 that use batched matmuls. Skip it
during reservation since the memory tracking is already handled by
the pool allocations.
Fixes#15249
* model/parsers: fix gemma4 arg parsing when quoted strings contain "
Fixes: #15241
* add more tests, be careful about what we escape
We want Windows-style paths to not get misinterpreted
* fix backslash-quote case, it really should be a literal backslash
h/t to @chathaway-codes for pointing this out!
Co-Authored-By: Charles H <2773397+chathaway-codes@users.noreply.github.com>
---------
Co-authored-by: Charles H <2773397+chathaway-codes@users.noreply.github.com>
Add --num-ctx flag to set context size, and report NumCtx in model info
header. Calibrate tokens-per-word ratio during warmup using actual
tokenization metrics from the model, replacing the fixed 1.3 heuristic.
This produces more accurate prompt token counts for --prompt-tokens.
Also add fetchContextLength() to query running model context via /api/ps.
* tokenizer: add byte fallback for SentencePiece BPE encoding
When BPE merging produces tokens not in the vocabulary, fall back to
encoding each UTF-8 byte as <0xHH> byte tokens instead of silently
dropping the character. Also teach Decode to convert <0xHH> tokens
back to raw bytes.
Fixes#15229, fixes#15231
* tokenizer fixes
* bench: add prompt calibration, context size flag, and NumCtx reporting
Add --num-ctx flag to set context size, and report NumCtx in model info
header. Calibrate tokens-per-word ratio during warmup using actual
tokenization metrics from the model, replacing the fixed 1.3 heuristic.
This produces more accurate prompt token counts for --prompt-tokens.
Also add fetchContextLength() to query running model context via /api/ps.
* integration: improve vision test robustness and add thinking tests
Add skipIfNoVisionOverride() to skip vision tests when OLLAMA_TEST_MODEL
is set to a non-vision model. Add Think:false to context exhaustion test
to prevent thinking models from using all context before the test can
measure it. Add third test image (ollama homepage) and replace OCR test
with ImageDescription test using it. Relax match strings for broader
model compatibility. Add TestThinkingEnabled and TestThinkingSuppressed
to verify thinking output and channel tag handling.
* gemma4: add Gemma 4 GGML model support
Add full Gemma 4 model family support (E2B, E4B, 26B MoE, 31B Dense)
for the GGML backend including text, vision, converter, parser, and
renderer.
Text model features:
- Sliding window + full attention with per-layer patterns
- KV sharing across layers with donor map
- Per-layer embeddings (PLE) with learned projections
- MoE routing with RMSNorm + learned scale
- Proportional RoPE with freq_factors for global attention
- Final logit softcapping
Vision model features:
- SigLIP vision encoder with 2D RoPE
- ClippableLinear with input/output clamping via packed v.clamp_data
- Adaptive average pooling with nMerge kernel
- Multi-modal projection with unweighted RMSNorm
Converter:
- Safetensors to GGUF with vision tensor renaming
- Fused MoE gate_up_proj splitting
- Vision patch embedding reshape (HF to Conv2D layout)
- Packed clamp data tensor for ClippableLinear bounds
- Proportional RoPE freq_factors generation
Also includes:
- BackendGet() on ml.Tensor for reading weight tensor data
- Q6_K CUDA get_rows kernel support
- MoE-aware ffn_down quantization layer counting
- Gemma4 parser with tool calling and thinking support
- Gemma4 renderer with structured tool format
- Architecture-based auto-detection of renderer/parser/stop tokens
- Integration test gemma4 model list additions
* gemma4: add audio support with USM conformer encoder
Add audio encoding for Gemma 4 using the USM conformer architecture:
- Converter: audio tensor mapping, SSCP/conformer/embedder name replacements,
softplus repacker for per_dim_scale, F32 enforcement for conv weights
- GGML backend: Conv1DDW and PadExt tensor ops
- Audio encoder: SSCP Conv2D, 12 conformer blocks (FFW + block-local
attention with relative position embeddings + LightConv1d + FFW),
output projection, audio-to-text embedding projector
- Audio preprocessing: WAV decode, mel spectrogram, FFT (pure Go)
- Model wiring: WAV detection, audio token handling, unified PostTokenize
Correctly transcribes "why is the sky blue" from test audio.
* integration: add gemma4 audio tests including OpenAI API coverage
Test audio transcription and response via the Ollama native API, plus
two new tests exercising the OpenAI-compatible endpoints:
- /v1/audio/transcriptions (multipart form upload)
- /v1/chat/completions with input_audio content type
All tests use capability checks and skip models without audio support.
* gemma4: add OpenAI audio API support and capability detection
- Add CapabilityAudio and detect from audio.block_count in GGUF
- Add /v1/audio/transcriptions endpoint with TranscriptionMiddleware
- Add input_audio content type support in /v1/chat/completions
- Add TranscriptionRequest/Response types in openai package
* gemma4: add audio input support for run command
- /audio toggle in interactive mode for voice chat
- Platform-specific microphone recording (AVFoundation on macOS,
PulseAudio/ALSA on Linux, WASAPI on Windows)
- Space to start/stop recording, automatic chunking for long audio
* gemma4: add transcribe command (ollama transcribe MODEL)
- Interactive mode with readline prompt and slash commands
- Non-interactive mode for piped audio or record-until-Ctrl+C
- Chunked streaming transcription for long recordings
- Word-wrapped output matching run command style
* gemma4: add parser, renderer, and integration test plumbing
* gemma4: fix renderer to emit BOS token
* gemma4: add OpenAI audio transcription API and input_audio support
* gemma4: update converter for new weight drop naming
* gemma4: add per_expert_scale to MoE router and fix moe_intermediate_size config
* gemma4: rewrite renderer to match HF Jinja2 template exactly
Fix 8 bugs found by building 55 reference tests verified against the
HF Jinja2 chat template (VERIFY_JINJA2=1 shells out to Python):
- Tool responses use separate <|turn>tool turns (not inline tags)
- Tool calls emitted before content in assistant messages
- Thinking content stripped from assistant history (strip_thinking)
- User, tool, and system content trimmed (template does | trim)
- Empty system message still emits system turn (check role, not content)
- Nested object properties rendered recursively with required field
- Array items specification rendered for array-type properties
- OBJECT/ARRAY type-specific rendering comma logic matches template
Also adds Required field to api.ToolProperty for nested object schemas,
replaces old gemma4_test.go with comprehensive gemma4_reference_test.go,
and commits the Jinja2 template as testdata for verification.
* gemma4: fix MoE fused gate_up split and multiline tool-call arg parsing
- Text MoE: split `ffn_gate_up_exps` into contiguous `[gate|up]` halves instead of stride-2 slices.
- Parser: escape control characters in `<|"|>...<|"|>` string literals when converting tool-call args to JSON.
- Fixes warnings like `invalid character '\n' in string literal` for multiline tool arguments.
- Add Gemma4 parser regressions for multiline tool-call args and `gemma4ArgsToJSON`.
* cmd: simplify audio input to dropped file attachments
* gemma4: use full SWA memory for better cache reuse
* gemma4: initialize clamps after backend load
* convert: align gemma4 audio tensor renames with llama.cpp
* Remove redundant comments in gemma4 vision model
* Format Gemma4 MoE block field alignment
* use 4096 kvcache.NewSWAMemCache
* convert: support new Gemma4 audio_tower tensor naming (#15221)
Co-authored-by: jmorganca <jmorganca@gmail.com>
* fix integration test defaults for audio
* review comments and lint fixes
* remove unused audio/video files
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
Previously we were accidentally using different clients/UAs depending on
whether it was an inference call or a different call. This change makes
them consistent, other than the timeout being different.
* tokenizer: add SentencePiece-style BPE support
Add WithSentencePieceNormalizer option to BytePairEncoding for models
that use BPE with SentencePiece-style space markers (space to/from
U+2581).
NewBytePairEncoding is unchanged; the new NewBytePairEncodingWithOptions
constructor accepts BPEOption functions. Decoding handles the reverse
mapping of U+2581 back to spaces.
* review comments
Replace hardcoded Encode(prompt, true) with
Encode(prompt, r.Tokenizer.AddBOS()) so the pipeline respects each
model's tokenizer configuration.
Models with add_bos_token=true (gemma3, llama): unchanged, tokenizer
still prepends BOS.
Models with bos_token=null (qwen3, qwen3.5): unchanged, the BOS
guard (vocab.BOS >= 0) already prevented prepending regardless of
the flag.
This aligns the pipeline with the /v1/tokenize endpoint which already
uses Tokenizer.AddBOS().
pullModelManifest unmarshals the registry response into a Go struct
then re-marshals with json.Marshal before writing to disk. When the
registry's JSON formatting or field ordering differs from Go's
output, the local SHA256 won't match the registry's
Ollama-Content-Digest header, causing false "out of date" warnings.
Preserve the raw bytes from the registry response and write them
directly to disk so the local manifest is byte-for-byte identical
to what the registry serves.
* anthropic: fix empty inputs in content blocks
When we switched to `api.ToolCallFunctionArguments`, `omitempty` stopped
doing what we were relying on it for before. This would cause non-tool
content blocks to have an `"input": {}` field, which doesn't match our
old behavior.
* use omitzero instead
The staleness check compared the local manifest digest (SHA256 of the
file on disk) against the registry's Ollama-Content-Digest header.
These never matched because PullModel re-serializes the manifest JSON
before writing, producing different bytes than the registry's original.
The fallback comparison (local modified_at vs upstream push time) was
also broken: the generated TypeScript Time class discards the actual
timestamp value, so Date parsing always produced NaN.
Fix by moving the staleness comparison server-side where we have
reliable access to both the local manifest file mtime and the upstream
push time. The /api/v1/model/upstream endpoint now returns a simple
`stale` boolean instead of raw digests for the frontend to compare.
Also adds User-Agent to the CORS allowed headers for dev mode.
A stop-gap for now to guide users better. We'll add more in-depth recommendations per integration as well.
---------
Co-authored-by: Parth Sareen <parth.sareen@ollama.com>
Add periodic snapshots every 8k tokens and near the end of the prompt
so that long prompts can be partially restored and thinking/generation
can be retried without full reprocessing.
Update LRU last used time just on the nodes that actually used
during processing rather than all snapshots along the path. This
allows eviction to remove nodes more accurately so we can avoid
other heuristics to auto-merge nodes.
mlx.Copy shares the backing buffer with its source (via
copy_shared_buffer) rather than allocating independent storage.
When used to snapshot a slice of the KV cache, the snapshot array
holds the entire original cache buffer alive through the shared
data pointer — even after eval detaches the computation graph.
Replace Copy with Contiguous in Snapshot and Split. Contiguous
allocates a compact buffer when the source buffer is significantly
larger than the logical slice (Contiguous::eval checks
buffer_size > nbytes + 16384), which is always the case for KV
cache slices.
Copilot Chat prefers to use `general.basename` in the built-in Ollama
integration, but this name isn't usually shown directly to users (and
there may be many models that share this name). Instead we pass back
`req.Model`, which for this extension is the value that we return from
`/api/tags`
