KV cache quantization has a dependency on the flash attention kernel.
We currently cannot use flash attention with gpt-oss as it requires
additional operations.
The model definition does not call flash attention, so it works
regardless of the setting but the cache will pick up the
quantization type. This updates the flash attention setting earlier
in the loading flow so that all downstream settings are also set correctly.
Fixes: #11671
* bf16
* tests
* gpt-oss
* enable gptoss for engine
* rough estimate
* convert to mxfp4
* handle safetensors U8
* clamp glu/linear
* update tokenizer
* MXFP4 support
This implements the Open Compute Microscaling (MX) FP4 format
as a tensor type with backend implementations focusing
on mulmat and mulmatid on CPU, CUDA, and Metal.
* Unit tests for MXFP4 support
This exercises various operations and shapes on both CPU and GPU (if detected
on the system)
* cuda graph
* unit test adjustments
* cuda: optimize memory access
Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4
* mac: fix crash on old macos versions
cblas_sgemm is only supported on v13.3 and up, however bf16 is
only supported on v14+ so we were falling back to ggml-blas and
crashing on bf16 tensors. Checking for the function being null
seems to be the simplest way to condittionally avoid registering the
backend.
* server: Minimum context length for gptoss
This model requires a minimum context length of 8192 to function
effectively. Users can set higher values through all normal mechanisms
but lower values will be silently reset.
* ggml: Multiply by numParallel for gptoss sliding window
When computing the graph size estimate, the context size is already
multiplied by numParallel so estimates reflect that. However, since
sliding window models use a smaller, fixed context size, they need
to manually take numParallel into account.
* gpt-oss integration
includes harmony parser and thinking levels, etc.
* fix sync
* fix tests
* fix lint
---------
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Jesse Gross <jesse@ollama.com>
Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
There is a bug when using sliding window attention where we run
out of KV cache slots. This is likely due to not correctly removing
all of the entries as they slide out of range. This adds additional
logging when this occurs to track down the source.
Bug #10127
Models that use sliding window attention can only resume a sequence
from the cache if it falls within the saved windows. This works well
if the next message picks up where the old one left off. However, it
generally prevents a partial prefix match unless the entire conversation
falls within the sliding window.
This can be a problem with reasoning models where the traces are
supposed to be removed from future messages, forcing the entire
history to be re-evaluated.
This change allows models to specify that a larger amount of the
history be retained in memory, to allow more partial resumption.
It still respects the window that the model was trained on for
token generation.
* Enable CUDA Graphs for gemma3n.
Similar to
https://github.com/ggml-org/llama.cpp/pull/14741,
though ollama has a slightly different model graph
than llama.cpp which requires different workaround
checks.
* Remove residual check by reshaping differently in gemma3n model
This should make the heuristics more robust
When we context shift, we delete half the context and apply RoPE
with an offset to the other half. We used to RoPE across the entire
context in a single pass with a zero offset for the deleted
section. With the change to shifting in batches, we can skip any
batches where all of the offsets would be zero. This typically
reduces the number of operations by half.
Currently, when we need to do a shift on the cache, it is one
RoPE operation on the entire size of the cache (per layer). In
some cases, this can create a compute graph that is larger than
the forward pass since the forward pass is working in batches.
Since we don't consider shifting in our memory estimates, it's
possible for this to cause a crash if we run out of memory.
By limiting the size of the RoPE calls to batch size chunks, we
ensure that the shift will never exceed the size of the forward
pass, since the forward pass will also contain a RoPE of the same
size. This does not have a sigificant impact on performance since
RoPE is a math operation that is mostly proportional to the size
of its inputs.
In theory defrag could have the same issue since it also creates a
compute graph outside of the forward pass, however, since it is
only copies, it does not require any working space.
StatusError was unreachable, the client always checked for error messages in the response body first, and the server always includes error messages with HTTP error status codes.
Reporting params.NumGPULayers can be misleading because it is the
requested number of layers, not the actual number that is loaded.
While they are often the same, there are cases where they might mismatch,
such as if the GPU backend is missing.
We don't get valid UUIDs for AMD GPUs on Windows, so the best option
is to use the ordinal IDs. This brings us in line with what we currently
do on the Ollama server - the only exception is AMD GPUs on Linux, which
falls back to using ordinal IDs. The GGML implementation has no fallback
but it doesn't appear to occur for any of the GPUs that we support.
It's also possible that there are collisions between ordinal IDs for
different libraries - however the only places where we use them are
AMD on Windows and Metal on Mac, which can never occur on the same
system.
The current scheduler algorithm of picking the paralellism based on available
VRAM complicates the upcoming dynamic layer memory allocation algorithm. This
changes the default to 1, with the intent going forward that parallelism is
explicit and will no longer be dynamically determined. Removal of the dynamic
logic will come in a follow up.
