Problem: Model loading takes 2-3 minutes on first load with no user feedback,
causing confusion about whether the system is frozen or working.
Root Cause: GPU initialization (reserveWorstCaseGraph) takes ~164 seconds on
Tesla K80 GPUs due to CUDA kernel compilation (PTX JIT for compute 3.7). This
is by design - it validates GPU compatibility before committing to full load.
Solution:
1. Add comprehensive timing instrumentation to identify bottlenecks
2. Add user-facing progress messages explaining the delay
Changes:
- cmd/cmd.go: Update spinner with informative message for users
- llama/llama.go: Add timing logs for CGO model loading
- runner/llamarunner/runner.go: Add detailed timing for llama runner
- runner/ollamarunner/runner.go: Add timing + stderr messages for new engine
- server/sched.go: Add timing for scheduler load operation
User Experience:
Before: Silent wait with blinking cursor for 2-3 minutes
After: Rotating spinner with message "loading model (may take 1-3 min on first load)"
Performance Metrics Captured:
- GGUF file reading: ~0.4s
- GPU kernel compilation: ~164s (bottleneck identified)
- Model weight loading: ~0.002s
- Total end-to-end: ~165s
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
This commit represents a complete rework after pulling the latest changes from
official ollama/ollama repository and re-applying Tesla K80 compatibility patches.
## Key Changes
### CUDA Compute Capability 3.7 Support (Tesla K80)
- Added sm_37 (compute 3.7) to CMAKE_CUDA_ARCHITECTURES in CMakeLists.txt
- Updated CMakePresets.json to include compute 3.7 in "CUDA 11" preset
- Using 37-virtual (PTX with JIT compilation) for maximum compatibility
### Legacy Toolchain Compatibility
- **NVIDIA Driver**: 470.256.02 (last version supporting Kepler/K80)
- **CUDA Version**: 11.4.4 (last CUDA 11.x supporting compute 3.7)
- **GCC Version**: 10.5.0 (required by CUDA 11.4 host_config.h)
### CPU Architecture Trade-offs
Due to GCC 10.5 limitation, sacrificed newer CPU optimizations:
- Alderlake CPU variant enabled WITHOUT AVX_VNNI (requires GCC 11+)
- Still supports: SSE4.2, AVX, F16C, AVX2, BMI2, FMA
- Performance impact: ~3-7% on newer CPUs (acceptable for K80 compatibility)
### Build System Updates
- Modified ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt for compute 3.7
- Added -Wno-deprecated-gpu-targets flag to suppress warnings
- Updated ml/backend/ggml/ggml/src/CMakeLists.txt for Alderlake without AVX_VNNI
### Upstream Sync
Merged latest llama.cpp changes including:
- Enhanced KV cache management with ISWA and hybrid memory support
- Improved multi-modal support (mtmd framework)
- New model architectures (Gemma3, Llama4, Qwen3, etc.)
- GPU backend improvements for CUDA, Metal, and ROCm
- Updated quantization support and GGUF format handling
### Documentation
- Updated CLAUDE.md with comprehensive build instructions
- Documented toolchain constraints and CPU architecture trade-offs
- Removed outdated CI/CD workflows (tesla-k80-*.yml)
- Cleaned up temporary development artifacts
## Rationale
This fork maintains Tesla K80 GPU support (compute 3.7) which was dropped in
official Ollama due to legacy driver/CUDA requirements. The toolchain constraint
creates a deadlock:
- K80 → Driver 470 → CUDA 11.4 → GCC 10 → No AVX_VNNI
We accept the loss of cutting-edge CPU optimizations to enable running modern
LLMs on legacy but still capable Tesla K80 hardware (12GB VRAM per GPU).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Resolves two critical issues preventing robust model switching:
1. Scheduler deadlock: Fixed improper loop control flow that prevented
model unloading from triggering after conflict detection. Added proper
multi-GPU conflict detection and unload sequencing.
2. Silent inference failures: Changed critical cudaSetDevice() calls from
graceful error handling back to CUDA_CHECK to prevent models from
appearing to load successfully but failing silently during inference.
Result: Robust Tesla K80 dual-GPU model switching with self-healing
recovery instead of requiring system reboots.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add Gemma3n model support with text generation capabilities
- Add new CUDA mean operations for improved performance
- Add macOS documentation and performance tests
- Update LLAMA patches for ROCm/CUDA compatibility
- Fix various model conversion and processing issues
- Update CI workflows and build configurations
- Add library model tests and Shakespeare test data
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
When the same model is being reloaded rapidly with client connections
being canceled before the model finishes loading, the queued unload
event could cause a leak of runners by deleting a different runner from
the loaded list.
If a model is loading, and the request context is canceled during the load
by a client closing the connection, and another request is inbound for the
same model with a different configuration (context size, etc.) thus requiring
a reload, two unload events can be in flight. The first shuts down the
original model load, but the second one caused the loss of the new
reloading runner reference, thus triggering the leak.
The primary fix is detecting the duplicate unload and ignoring the second
instance. The load routine is also hardened to ensure we detect
clobbering an already present runner and unload it with a warning.
This enhances our logging in the scheduler. The initial "waiting for server" log
no longer claims an initial error state (now "not responding" which better reflects
the actual state). Runners now have slog wiring to report more details about the
runner, including PID.
* Adjust initial scheduler refCount
Ensure we only set the refCount on success
* sched: fix lock order inversion deadlock
Under certain race conditions, there was a scenario where the scheduler would
get into a deadlock while trying to update free space information while a model
was trying to unload.
this is in part to "pay" for #10452, which doubled the default context length. The combination isn't fully neutral though, because even though the old 4x2k limit and the new 2x4k limit are memory equivalent, the 1x fallback is larger with 4k
* increase default context length to 4096
We lower the default numParallel from 4 to 2 and use these "savings" to
double the default context length from 2048 to 4096.
We're memory neutral in cases when we previously would've used
numParallel == 4, but we add the following mitigation to handle some
cases where we would have previously fallen back to 1x2048 due to low
VRAM: we decide between 2048 and 4096 using a runtime check, choosing
2048 if we're on a one GPU system with total VRAM of <= 4 GB. We
purposefully don't check the available VRAM because we don't want the
context window size to change unexpectedly based on the available VRAM.
We plan on making the default even larger, but this is a relatively
low-risk change we can make to quickly double it.
* fix tests
add an explicit context length so they don't get truncated. The code
that converts -1 from being a signal for doing a runtime check isn't
running as part of these tests.
* tweak small gpu message
* clarify context length default
also make it actually show up in `ollama serve --help`
With support for multimodal models becoming more varied and common it is important for clients to be able to easily see what capabilities a model has. Retuning these from the show endpoint will allow clients to easily see what a model can do.
Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.
Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.
This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.
Fixes#9730Fixes#9890
feat: add new Ollama engine using ggml through cgo
This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.
- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations
This is the first implementation of the new engine. Follow up PRs will implement more features:
- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
The Go runner does not have a problem with supporting parallel
requests for most multimodal models. Now that we won't be potentially
falling back to server.cpp, this restriction can be lifted.
However, the new mllama model can't support parallel requests, so we
will need to keep a restriction for that.
* Fix embeddings memory corruption
The patch was leading to a buffer overrun corruption. Once removed though, parallism
in server.cpp lead to hitting an assert due to slot/seq IDs being >= token count. To
work around this, only use slot 0 for embeddings.
* Fix embed integration test assumption
The token eval count has changed with recent llama.cpp bumps (0.3.5+)
In mult-brand GPU setups, if we couldn't fully load the model we
would fall through the scheduler and mistakenly try to load across
a mix of brands. This makes sure we find the set of GPU(s) that
best fit for the partial load.
This change fixes the handling of keep_alive so that if client
request omits the setting, we only set this on initial load. Once
the model is loaded, if new requests leave this unset, we'll keep
whatever keep_alive was there.
Users may not realize the siny new model they're trying to load
fits on their disk, but can't load into system+GPU memory. Today
we crash, but with this fix, we'll give them a better error message
before even trying to load it.
Previously, some costly things were causing the loading of GGUF files
and their metadata and tensor information to be VERY slow:
* Too many allocations when decoding strings
* Hitting disk for each read of each key and value, resulting in a
not-okay amount of syscalls/disk I/O.
The show API is now down to 33ms from 800ms+ for llama3 on a macbook pro
m3.
This commit also prevents collecting large arrays of values when
decoding GGUFs (if desired). When such keys are encountered, their
values are null, and are encoded as such in JSON.
Also, this fixes a broken test that was not encoding valid GGUF.
Until ROCm v6.2 ships, we wont be able to get accurate free memory
reporting on windows, which makes automatic concurrency too risky.
Users can still opt-in but will need to pay attention to model sizes otherwise they may thrash/page VRAM or cause OOM crashes.
All other platforms and GPUs have accurate VRAM reporting wired
up now, so we can turn on concurrency by default.
This adjusts our default settings to enable multiple models and parallel
requests to a single model. Users can still override these by the same
env var settings as before. Parallel has a direct impact on
num_ctx, which in turn can have a significant impact on small VRAM GPUs
so this change also refines the algorithm so that when parallel is not
explicitly set by the user, we try to find a reasonable default that fits
the model on their GPU(s). As before, multiple models will only load
concurrently if they fully fit in VRAM.
While models are loading, the VRAM metrics are dynamic, so try
to load on a GPU that doesn't have a model actively loading, or wait
to avoid races that lead to OOMs
