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 fixes the issue where large models (>10B parameters) fail to
load due to underestimated compute buffer memory requirements, causing
allocation failures when the model should use multiple GPUs.
Problem:
- deepseek-r1:14b (14B, qwen2 architecture) failed with "failed to allocate
compute buffers" error
- System has 2×Tesla K80 GPUs (24GB total) but tried to fit 12GB model in
1×11GB GPU
- Root cause: Memory estimation underestimated compute buffers by 3-4×
(estimated 916 MB, actual requirement ~3-4 GB)
Solution:
1. Added model-family-specific batch size defaults (llm/memory.go)
- Different architectures have different optimal batch sizes
- deepseek2: 2048/256, qwen2: 512/512, llama: 512/512, etc.
- Ensures accurate memory estimation based on architecture
2. Updated server to use architecture-specific batch sizes (llm/server.go)
- Detects model architecture from GGUF metadata
- Uses family defaults when user doesn't specify
- Ensures consistency between estimation and allocation
3. Applied 3.5× safety margin to compute buffer estimates (llm/memory.go)
- Accounts for temporary tensors not captured in GraphSize formulas
- Conservative approach prevents allocation failures
- Documented with detailed analysis of underestimation causes
4. Implemented measurement API for future use (llama-context.cpp, llama.go)
- C++ function to measure actual memory requirements
- Go wrapper for integration into GPU selection
- Foundation for future measurement-based approach
- Currently unused but documented for future improvement
Results:
- deepseek-r1:14b now loads successfully using both GPUs
- Proper distribution: 25 layers on GPU0, 24 layers on GPU1
- Total memory: 16.2 GB across 2×11 GB GPUs (8.4 + 7.8 GB)
- Compute buffers: 3.1 GB per GPU (with safety margin applied)
- All other models continue to work correctly
Comprehensive documentation added to all modified code explaining:
- Problem analysis with real examples
- Solution rationale and trade-offs
- Future improvement paths
Tested with: deepseek-r1:14b, deepseek-r1:8b, gemma3:4b, gpt-oss
🤖 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>
* get eos_token_id from generation_config.json
* refactor
* include both ids and strings in trace
* comments
* remove special case for gemma3 special vocab (#10743)
* Move quantization logic to GGML via new backend
This moves the model aware logic to Go code and calls GGMLs quantization code for model creation.
* Remove "add model quantizations"
This is no longer needed now that quantization is implemented in Go+GGML code directly.
Some options listed in api/types.go are not supported in
newer models, or have been deprecated in the past. This is
the first of a series of PRs to clean up the API options
Clear KV cache when shift operation is not supported by model.
Added KvCacheCanShift() check to handle models that can't perform cache shifts,
falling back to full cache clear while preserving logical token history to
maintain expected behavior when context window fills up.
- output backend system info when initializing the backend. this ensures
this information is always present without needing to be called
explicitly
- convert to structured logging
- enumerate devices rather than backends since devices are ordered
- track device indices grouped by device name
Shield the code processing the embedding result
from subsequent calls that may overwrite the same
buffer to process a second input when retrieving
model embeddings.
* add build to .dockerignore
* test: only build one arch
* add build to .gitignore
* fix ccache path
* filter amdgpu targets
* only filter if autodetecting
* Don't clobber gpu list for default runner
This ensures the GPU specific environment variables are set properly
* explicitly set CXX compiler for HIP
* Update build_windows.ps1
This isn't complete, but is close. Dependencies are missing, and it only builds the "default" preset.
* build: add ollama subdir
* add .git to .dockerignore
* docs: update development.md
* update build_darwin.sh
* remove unused scripts
* llm: add cwd and build/lib/ollama to library paths
* default DYLD_LIBRARY_PATH to LD_LIBRARY_PATH in runner on macOS
* add additional cmake output vars for msvc
* interim edits to make server detection logic work with dll directories like lib/ollama/cuda_v12
* remove unncessary filepath.Dir, cleanup
* add hardware-specific directory to path
* use absolute server path
* build: linux arm
* cmake install targets
* remove unused files
* ml: visit each library path once
* build: skip cpu variants on arm
* build: install cpu targets
* build: fix workflow
* shorter names
* fix rocblas install
* docs: clean up development.md
* consistent build dir removal in development.md
* silence -Wimplicit-function-declaration build warnings in ggml-cpu
* update readme
* update development readme
* llm: update library lookup logic now that there is one runner (#8587)
* tweak development.md
* update docs
* add windows cuda/rocm tests
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Previously we decoded and re-encoded JSON schemas during validation,
which served no purpose since json.RawMessage already validates JSON
syntax. Worse, the re-encoding lost field ordering from the original
schema, which affects inference quality during step-by-step reasoning.
While fixing this ordering issue by using json.RawMessage directly,
testing revealed that schema_to_grammar (from llama.cpp) also fails to
preserve field order during grammar generation. This appears to be the
root cause of inference degradation.
This change prevents us from mangling the user's original schema order,
but we still need to address the ordering issue in schema_to_grammar.
That will be a separate change.
Updates #7978
The final implementation of #7499 removed dynamic vector requirements
in favor of a simpler filename based model, and this was left over logic that
is no longer needed.
* llama: wire up builtin runner
This adds a new entrypoint into the ollama CLI to run the cgo built runner.
On Mac arm64, this will have GPU support, but on all other platforms it will
be the lowest common denominator CPU build. After we fully transition
to the new Go runners more tech-debt can be removed and we can stop building
the "default" runner via make and rely on the builtin always.
* build: Make target improvements
Add a few new targets and help for building locally.
This also adjusts the runner lookup to favor local builds, then
runners relative to the executable, and finally payloads.
* Support customized CPU flags for runners
This implements a simplified custom CPU flags pattern for the runners.
When built without overrides, the runner name contains the vector flag
we check for (AVX) to ensure we don't try to run on unsupported systems
and crash. If the user builds a customized set, we omit the naming
scheme and don't check for compatibility. This avoids checking
requirements at runtime, so that logic has been removed as well. This
can be used to build GPU runners with no vector flags, or CPU/GPU
runners with additional flags (e.g. AVX512) enabled.
* Use relative paths
If the user checks out the repo in a path that contains spaces, make gets
really confused so use relative paths for everything in-repo to avoid breakage.
* Remove payloads from main binary
* install: clean up prior libraries
This removes support for v0.3.6 and older versions (before the tar bundle)
and ensures we clean up prior libraries before extracting the bundle(s).
Without this change, runners and dependent libraries could leak when we
update and lead to subtle runtime errors.
Fragmentation of the KV cache can occur due to cache shifting or
different sequences getting processed. Decode uses a heuristic to
decide if it should defrag. However, this heuristic isn't 100%
accurate, so decoding can sometimes fail by surprise.
For these cases, if decode indicates that there is no KV cache space,
we should defrag and then try again.
Check for NULL return values from llama.cpp in more places and
convert them into Go errors, which should make debugging easier
in the future rather than having hidden surprises in our data
structures.
Mllama has large embeddings (100 MB per image) and each embedding is
represented as 1 token when passed to llama.cpp. Batches are pre-
allocated for the size of the tokens times the batch size, so this
results in allocations of over 50 GB at the default batch size.
On some systems, these mallocs will fail.
Since an image is represented as a single token and mllama doesn't
support more than 1 image per request, we only need to allocate a
batch size of 1, which is much more reasonable. In addition, for
non-multimodal models, we don't need to allocate the embedding
batches at all.
Fixes#7464
-Update mllama to take the cross attention state as embeddings in
a batch, more similar to how Llava handles it. This improves
integration with the input cache.
-Pass locations in a prompt for embeddings using tags similar to Llava.
-Abstract interface to vision models so the main runner accesses Clip
and Mllama similarly
Co-authored-by: Michael Yang <mxyng@pm.me>
This will no longer error if built with regular gcc on windows. To help
triage issues that may come in related to different compilers, the runner now
reports the compier used by cgo.
Llama.cpp sometimes returns NULL as a return value to report an
error. We should explicitly check for this and convert it to a Go
error rather than putting NULL in our data structures and waiting
for it to blow up later.
* fix(ext_server): Port llama.cpp sampling refactors to ext_server
This was a fairly large changeset. I closely followed the changes here:
df270ef745
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(server.cpp): Refactor server.cpp logging for llama.cpp overhaul
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Bump llama.cpp to the latest master with `granite` support
This does not yet have granite MoE support, but that can come in a
follow up PR
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(patches): Update all patches (except solar-pro) to work with bumped llama.cpp
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(solar): Update solar patch for llama.cpp bump
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama.cpp): Bump llama.cpp for granitemoe support
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama.cpp): Bump llama.cpp for granitemoe support
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(solar): Update the solar-pro patch for latest llama.cpp bump
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama.cpp): Bump to the latest master of llama.cpp
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(patches): Update all patches for latest bump
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama): Always run sync.sh from the right directory
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama/patches): Update llama patches
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat(llama)!: Rough sync with llama.cpp submodule
There are a number of changes that will need to be propagated to llama.go
before any of this works!
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama/patches): Add a patch and update for missing ggml-impl.h include
This include is where the ggml_cgraph struct is defined. It is included in
many of the .c files to define the forward declartion in ggml.h. It seems
that with the subset of code included here, the import was somehow lost (or
out-of-order) when building, so adding this include to llama.cpp fixes the
missing definition.
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama/sync): Add missing ggml-cpu-impl.h copy-over in sync.sh
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Add missing log.cpp
This was added as part of the logging overhaul done in llama.cpp
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Overhaul use of sampling module for llama.cpp changes
The changes here reflect the changes made in the big llama.cpp sampling PR
https://github.com/ggerganov/llama.cpp/pull/9294
The sampling functionality is now broken into the base interface
(llama_sampler) and the generation implementation (gpt_sampler). The
changes here reflect that. Since the sampling.h/sampling.cpp code uses c++
STL headers, the sampling_ext.[h|cpp] wrapper is maintained to allow go to
access a pure-C interface.
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Fix the impl of SampleTokenGreedy for new sampling
I don't think this method is currently used, so it could probably just be
removed so that all sampling goes through the GPT interface, but in the
interest of doing no harm, this should keep the method working as expected.
Branch: IBMGraniteArchitectureSupport
* fix(llama): Remove unused SampleTokenGreedy
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(sync): Remove bash-specific change to sync.sh
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* chore(gofumpt): Format on llama.go to pass linting
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llm): Fix missing <thread> include in ext_server
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Remove TODO about grammar_first
This feature was not used/needed previously so should be fine without
plumbing it through now.
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Better naming for sampling wrapper and args
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Fix patch 05 to use new wrapper api and re-sync
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* runner: Flush pending responses before returning
If there are any pending reponses (such as from potential stop
tokens) then we should send them back before ending the sequence.
Otherwise, we can be missing tokens at the end of a response.
Fixes#6707
* fix(llama/sampling): Use gpt_sampler with a forward declaration
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llama): Remove unnecessary patch for gguf impl header
This was caused by an earlier mistake in the embeddings patch that was
dereferencing the pointer instead of using the wrapper API.
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(llm): Remove use of deprecated --log-disable flag
Branch: IBMGraniteArchitectureSupport
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Having v11 support hard-coded into the cgo settings causes warnings
for newer Xcode versions. This should help keep the build clean for users
building from source with the latest tools, while still allow us to target
the older OS via our CI processes.
