Use nthreads=128 for ncols=4 configurations in flash attention tile
kernel to reduce shared memory usage below 48KB limit on Maxwell
architectures (sm_50/52).
With nthreads=256 and ncols=4, np=2 which caused shared memory to
exceed 48KB. With nthreads=128 and ncols=4, np=1 keeps shared memory
under the limit.
The nvidia_fp32 config for (576, 512) head sizes had nbatch_fa=32,
which caused zero-sized arrays when computing array dimensions:
nbatch_fa / (np * warp_size) = 32 / (2 * 32) = 0
This resulted in CUDA compilation failures on CUDA 12 (Windows and
Linux arm64):
- "static assertion failed with nbatch_fa % (np*warp_size) != 0"
- "the size of an array must be greater than zero"
Fix by changing nbatch_fa from 32 to 64 for all (576, 512) configs
in the nvidia_fp32 function, matching the nvidia_fp16 and AMD configs.
* model: add MLA absorption for glm4moelite
Split the combined KV_B tensor into separate K_B and V_B tensors
during conversion, enabling MLA (Multi-head Latent Attention)
absorption which compresses the KV cache for improved efficiency.
* ggml: enable MLA flash attention for GLM-4.7-flash
Add support for gqa_ratio 4 in MLA flash attention kernels. GLM-4.7-flash
uses head size 576 with gqa_ratio 4, which was previously only supported
for gqa_ratio 16 (DeepSeek).
Metal changes:
- Enable head size 576 for flash attention
- Increase simdgroups to 8 for large heads (>=512)
- Add case 8 kernel dispatch for 8 simdgroups
CUDA changes:
- Add gqa_ratio 4 support for head 576/512
- Add tile configs for (576, 512, 4) and (576, 512, 8)
- Add MMA config cases for ncols 4
- Add template instances for ncols2=4
* model: add compatibility validation for glm4moelite architecture
* Revert "add support for NVIDIA Nemotron 3 Nano"
This reverts commit e7d2ae9d69.
* GGML update to 380b4c984
Remove MaskBatchPadding as GGML_KQ_MASK_PAD is no longer present (no
padding required)
* update to c45f89d55
* ec98e2002
solar pro needed more adjusting - needs verification
* review comments
* flash attn: add auto mode for llama engine
If the user does not specify fa in the environment, use auto-mode.
* review comments
* ensure kv cache quantized types have FA explicitly enabled
additional review comments
* feat: Bump llama.cpp to the latest master (17f7f4b)
This brings in significant improvements to prefill performance for all
models using the SSM_CONV and SSM_SCAN ops (granite4, jamba, falcon-h,
nemotron-h, Qwen3 Next) on Apple Metal.
See https://github.com/ggml-org/llama.cpp/pull/17876
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 1-4
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Update patches 5-12
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 13-18
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patch 20
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Update patches 21-31
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Sync vendored code
The two files I'm not sure about here are the swap from gemma3-iswa.cpp to
gemma3.cpp (I chose to include this because I think it's required), and the
inclusion of `ggml-zendnn.h` which I chose to omit.
Branch: LlamaCPPMetalSSMImprovements
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Although the vision component of multimodal models typically already
call the optimized nn.Attention, it is converted into non-fused
operations. That is because the backend-specific fused kernels may
have requirements, such as padding, and they is performed by the
cache, which vision encoders don't use.
This implements a fallback path in the backend, softening the
requirements into optimizations. In turn, this allows flash attention
to be used for vision encoders, saving a significant amount of VRAM
and improving performance.
We currently use cache padding of 32 when not using flash attention
and 256 with flash attention, which is based on the historic alignment
requirements of these kernels. The restrictions have since been
loosened but there are still performance benefits, such as better
CUDA graph reuse.
Since the requirement is no longer kernel-specific, set the padding
uniformly to 256, as llama.cpp has.
* Revert "vulkan: temporary cary of vulkan fixes (#12971)"
This reverts commit 3a9e8e9fd4.
* ggml update to b7087
* fix argsort on metal
* update to b7108
* fix bakllava regression
This model lacks the metadata for the projector type.
* update to b7209
* fix TopK perf
* only build arm code on arm
We now do a deeper probe of CUDA devices to verify the library version has
the correct compute capability coverage for the device. Due to ROCm also
interpreting the CUDA env var to filter AMD devices, we try to avoid setting
it which leads to problems in mixed vendor systems. However without setting
it for this deeper probe, each CUDA library subprocess discovers all CUDA GPUs
and on systems with lots of GPUs, this can lead to hitting timeouts. The fix is
to turn on the CUDA visibility env var just for this deeper probe use-case.
We currently copy data into the KV cache in contiguous buffers using
ggml_cpy(). ggml_set_rows() was introduced to allow scatter operation
so that contiguous buffers are no longer required. The direct primary
benefit of this is that we no longer need to perform defragmentation.
However, GGML recently removed an optimization for ggml_cpy() and
we picked it up in 544b673 "ggml update to b6840 (#12791)". This
caused a roughly 40% drop in token generation performance on CUDA
due to CUDA graphs no longer being used. By switching to
ggml_set_rows(), the original optimization is no longer necessary
and CUDA performance is restored.
Fixes#13112
GGML requires tensors to be contiguous for reshape and if
this is not the case, it will assert fail. Contiguous is an
expensive operation, so it's best to do it lazily when it is
actually required rather than ahead of time when it may not
be needed.
Calling abort on windows triggers the C++ runtime to attempt a debugger
attach, which causes the crashed runners to hang instead of exit, leading
to a timeout instead of a fast failure during discovery.
* build: optimize dockerfile context for iterating
This moves the copy of the source into the layer AFTER
doing software installs so we don't have to go through
the RPM install for cuda, etc. every time you touch a
source file.
* amd: implement linux sysfs based VRAM lookup
This adds a C++ implementation of sysfs DRM VRAM discovery
for more accurate free VRAM data on linux for AMD GPUs.
We currently assign model layers to GPUs according to free VRAM,
which assumes that GPU performance is roughly equal. This does not
work well for mixed dGPU and iGPU systems because iGPUs typically
use system memory which is large but their performance is slow.
This instead assigns layers to dGPUs first and then iGPUs.
In the future, this could be generalized to have a more fine grained
notion of GPU performance but dGPU vs. iGPU performance is the most
extreme.
We used to control the way that llama.cpp saw devices using
CUDA_VISIBLE_DEVICES or similar. This would ensure that the layers
offloaded to a device were actually the ones intended. This is
particularly important because we might reorder devices based on
free memory or performance.
When we started explicitly scheduling layers, this logic went
away but the llamarunner didn't have any way to set the correct
order of devices. This meant that the correct number of layers
would be assigned to a device but not necessarily the layers
that were expected. This change sets up the devices correctly
based on the offload information.
* discovery: only retry AMD GPUs
CUDA and Vulkan don't crash on unsupported devices, so retry isn't necessary.
This also refactors the code to shift the Library specific logic into the ml
package.
* review comments
* PDH free memory skeleton
* Add PDH printing
* Add LUID support for Vulkan
* wire luid from ggml-vulkan to mem-dxgi-pdh file
* Fix to ggml-impl
* Continue skeleton
* Implemented ggml_dxgi_pdh_get_device_memory
* fix comments
* Fix - change value GB to bytes
* add ifdefs to only support windows and not linux
* modify error codes
* Finished ggml_dxgi_pdh_init() function
* completed ggml_dxgi_pdh_release()
* Formatting changes, add static to functions
* fix build errors
* fix go build error
* fix luid - now should match between dxgi and vulkan
* Fix the free memory reporting (was using copy by value, change to reference)
* keep only dxgi1_2.h
* Modifications based on PR feedback
* fix merge conflicts (2) and fix desc1.description printout
* move dxgi + pdh api calls to before the vendor specific library calls
* change from 3 samples to 1 sample for PDH
* modify when old_mode is set
* add fix for building MacOS
* fix release and returns for other vendors
* add patch file
The initial implementation of qwen3-vl:235b exceeded the maximum graph
size based on the number of tensors. Although this was later fixed
through the use of the mrope operation, we are close to the limit in
some cases. This updates to track the current llama.cpp usage of GGML.
We pass invalid pointers when we check the size of the required
compute graph before fitting. Some CUDA APIs validate these pointers
but we can just skip them during this phase. cudaMemsetAsync is one
of these that we weren't skipping but never took the code path that
used it. Now that we have enabled op_offload, we can hit it in
memory pressured situations.
When a model is partially offloaded to system RAM, we can either
do the calculations on the CPU or we can temporarily transfer the
data to the GPU to do the calculations there. Small batches tend
to be better on the CPU, large batches on the GPU.
The llamarunner used the GPU in most cases and the ollamarunner
used the CPU. Although the ollamarunner saw an improvement in
token generation performance, there was a large performance hit
in prompt processing (3-10x).
There is an existing heuristic to dynamically switch between these
two modes but in practice it doesn't have enough information to
accurately make that decision. This adds authoritative data to make
the check work to get the best of both worlds.
Fixes#12037
* Fix vulkan PCI ID and ID handling
Intel GPUs may not report PCI IDs which was leading to incorrect overlap
detection. Switch to using the existing PCI IDs, however AMD GPUs claim not to
report PCI IDs, but actually do, so try anyway, as this is required for ADLX to
find the GPUs on Windows. Numeric IDs lead to scheduling problems, so this also
switches Vulkan to use UUID based IDs. The GPU discovery patches have been
squashed into a single patch to simplify future rebases.
* review comments
* DRY out the runner lifecycle code
Now that discovery uses the runners as well, this unifies the runner spawning code
into a single place. This also unifies GPU discovery types with the newer ml.DeviceInfo
* win: make incremental builds better
Place build artifacts in discrete directories so incremental builds don't have to start fresh
* Adjust sort order to consider iGPUs
* handle cpu inference oom scenarios
* review comments
Users on Windows without GPUs are reporting errors relating to
cudaDriverGetVersion with the device set to -1. This ensures we only grab the
driver once we're enumerating actual devices.
