* vulkan: Support asymmetric FA in coopmat2 path
There has been some recent interest/experimentation with mixed quantization
types for FA. I had originally designed the cm2 FA shader with this in mind
(because I didn't realize it wasn't supported at the time!), this change
adds the missing pieces and enables it.
Also support Q1_0 since people have been trying that out (seems crazy, but
who knows).
We should be able to do similar things in the coopmat1/scalar path, but
there's another change open against the scalar path and I don't want to
conflict.
* reorder cases
* vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it
* use FetchContent to get SPIRV-Headers
* Fetch spirv-headers unconditionally
* remove fetchcontent, rely on installed headers
* fix ubuntu job
* Update docs/build.md
* vulkan: change gated_delta_net to shard a column across a subgroup
This is based on https://github.com/ggml-org/llama.cpp/pull/20391, I used an
LLM to port the CUDA code to Vulkan, and guided to it to make various fixes to
work with Vulkan (e.g. handling different subgroup sizes, unknown mapping of
subgroup to invocation id, using subgroupAdd optionally, etc.).
This fixes a perf regression from the transposing of the values in memory
(!20443).
* vulkan: Spread columns across fewer lanes to reduce the number of workgroups
* vulkan: Fix data races in coopmat1 mul_mat(_id)
Add barriers between coopmat store and regular loads. We sort of got away with
this because it was the same subgroup accessing the values, but it's still a
race and may not work.
* switch to subgroup control barriers
* vulkan: split mul_mat into multiple dispatches to avoid overflow
The batch dimensions can be greater than the max workgroup count limit,
in which case we need to split into multiple dispatches and pass the base
index through a push constant.
Fall back for the less common p021 and nc variants.
* address feedback
The cpu and cuda backends use fp16 for the VKQ accumulator type, this change
does the same for vulkan. This helps particularly with large head sizes which
are very register-limited.
I tried this for the coopmat1 path and it slowed down a bit. I didn't try for
scalar.
I applied the softmax bias that the cuda backend uses to avoid overflow,
although I was not able to reproduce the original bug without it.
Write out a 2-bit code per block and avoid loading the mask when it
matches these two common cases.
Apply this optimization when the mask is relatively large (i.e. prompt
processing).
* vulkan: Remove transfer_ctx, do everything in compute_ctx.
We had a bug where a set_tensor_async (using transfer_ctx) didn't get
submitted before the graph_compute (using compute_ctx) that came after
it. To avoid this sort of issue, just do everything in compute_ctx.
Remove transfer_cmd_pool, which was already unused.
* fix crash with perf logger
Change ggml_vk_mul_mat_vec_id_q_f16 to loop over the batch dimension and
update the indexing calculations in get_offsets.
Mat-vec is faster than mat-mat for small values of n. We don't get the same
reuse of the weights as in the non-ID path, but with this the cost is linear
in n rather than n>1 being far slower than n==1.
This fixes incoherent output in Llama-4-Maverick-17B-128E-PAB-Q8_0, which
has a mul_mat_id with an A matrix that's Q8_0 8192 x 5120 x 128.
This should work when the number of blocks in the A matrix is less than 2^32
(for mul_mat_vec or mul_mm_cm2), or for mul_mm I think the limit is like
2^32*LOAD_VEC_A elements.
- Divide batch_stride by QUANT_K earlier, so the block index calculation works in 32b.
- Each vk_pipeline_struct has a linked list of pipelines that will allow it to handle
variants. So far this change just adds a single use case for this, compiling with the
e64BitIndexingEXT flag.
- Use the 64b indexing variant when the A matrix is larger than maxStorageBufferRange.
64-bit indexing has some cost - around 3-5% in MoE models, so it's worth the effort
to avoid enabling it unconditionally.
* vulkan: support buffer_from_host_ptr
* hacky use of buffer_from_host_ptr for directio
* disable buffer_from_host_ptr cap
* use external memory for ggml_vk_host_malloc, revert model loader changes
* disable external_memory_host for MoltenVK
* take buffer memory types into account
* don't use external_memory_host for ggml_vk_host_malloc
* vulkan: Optimize GGML_OP_CUMSUM
There are two paths: The preexisting one that does a whole row per workgroup
in a single shader, and one that splits each row into multiple blocks and does
two passes. The first pass computes partials within a block, the second adds
the block partials to compute the final result. The multipass shader is used
when there are a small number of large rows.
In the whole-row shader, handle multiple elements per invocation.
* use 2 ELEM_PER_THREAD for AMD/Intel
* address feedback
* vulkan: extend topk_moe to handle sigmoid w/exp_probs_b for nemotron
Also handle GGML_OP_SCALE at the end (nemotron, deepseek2).
Fewer pipeline variants and spec constants, just use push constants.
In test_topk_moe, change exp_probs_b to be 1D, matching real networks.
Update test-backend-ops and ggml-backend to allow verifying multiple outputs
in a fusion test (topk_moe has two outputs). Previously only the final node
was verified.
* change test_topk_moe to allow results in arbitrary order
* disable sigmoid fusion for moltenvk
* vulkan: Use BK=32 for coopmat2 mul_mat_id
* vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader
Disable robustness, remove the OOB check in decodeFuncB, and initialize the
row_ids to zero to avoid OOB access.
Don't slice/offset the B matrix to ic * BN, only to adjust the coord back down
to the range [0, BN) in decodeFuncB. Instead just slice with a row offset of
zero and remove the '& (BN - 1)'. This allows the compiler to common some of
the shared memory loads.
The goal is to enable the async loading code paths in
llama_model_loader::load_all_data, originally from #7896. This works and the
loads themselves are faster, but with host visible vidmem I think the cost of
allocating/mapping vidmem moves and becomes more expensive, and I don't see a
benefit by default. But with GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1 I do see a
significant improvement in model loading time.
I updated test_topk_moe to more closely match llm_graph_context::build_moe_ffn
and added coverage for exp_probs_b and some other missing combinations. This
exposed a bug in both CUDA and Vulkan backends where they were assuming the
input to argsort and the input to get_rows are the same. I'd like to optimize
this graph in another change, but for now just get it functional.
CUDA also had a bug where it got n_experts from the wrong place, leading to
GGML_ASSERT failures in some of the new tests.
This implements a variation of the perf logger where rather than timing each
operation individually with effectively a barrier in between, we put the
timing boundaries where we already synchronize and time the groups of work
that normally overlap. This can be useful to help understand whether
individual operations need to be optimized, or if the group is already running
efficiently.
GGML_VK_PERF_LOGGER_CONCURRENT=1 enables the new mode (when
GGML_VK_PERF_LOGGER is also set).
GGML_VK_SYNC_LOGGER=1 replaces the ENABLE_SYNC_LOGGING compile time switch.
When the number of cols is large, split each row across multiple workgroups.
There are three phases that communicate partial results through temp buffers:
(1) compute max partials
(2) take max of partials, compute sum(exp(x-max)) partials
(3) sum partials, compute scaled result
* vulkan: perf_logger improvements
- Move perf_logger from device to ctx.
- Add an env var to control the frequency we dump the stats. If you set a very
large value, it just dumps when the ctx is destroyed.
- Add a fusion info string to the tracking, only log one item per fused op.
- Fix MUL_MAT_ID flops calculation.
* fix vector sizes
The MoE models have a mul_mat_vec with very small m (32, 64, 128) right before
the topk_moe selection. Running multiple rows per wg doesn't utilize the SMs
well. I think even for larger m, f32 is so bandwidth-limited that running
multiple rows doesn't help.
* vulkan: Reduce temporary memory usage for TOP_K
- Compute row size for the temp buffer based on the output of the first pass.
- Update shader addressing math to use the output row size
- Pass the output row size as "ncols_output", what used to be "ncols_output" is now "k"
For the common case of K=40 and src0=(200000,1,1,1), this reduces the temporary buffer
from about 3.2MB to 500KB.
* vulkan: fix top_k bug when there are ties in the input
I noticed by inspection a bug in the vulkan top_k shader where if the least
value in the top_k appears multiple times we could end up writing those extra
copies out rather than some larger values (if the larger values are on higher
numbered threads).
I rewrote the test verification to handle this case, where the final index set
is not necessarily the same.
* Update tests/test-backend-ops.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
- Compute row size for the temp buffer based on the output of the first pass.
- Update shader addressing math to use the output row size
- Pass the output row size as "ncols_output", what used to be "ncols_output" is now "k"
For the common case of K=40 and src0=(200000,1,1,1), this reduces the temporary buffer
from about 3.2MB to 500KB.
* vulkan: Implement top-k
Each pass launches workgroups that each sort 2^N elements (where N is usually 7-10)
and discards all but the top K. Repeat until only K are left. And there's a fast
path when K==1 to just find the max value rather than sorting.
* fix pipeline selection
* vulkan: Add N-ary search algorithm for topk
* microoptimizations
* vulkan: support larger argsort
This is an extension of the original bitonic sorting shader that puts the
temporary values in global memory and when more than 1024 threads are needed
it runs multiple workgroups and synchronizes through a pipelinebarrier.
To improve the memory access pattern, a copy of the float value is kept with
the index value. I've applied this same change to the original shared memory
version of the shader, which is still used when ncols <= 1024.
* Reduce the number of shader variants. Use smaller workgroups when doing a single pass, for a modest perf boost
* reduce loop overhead
* run multiple cols per invocation, to reduce barrier overhead
* vulkan: change graph_compute to be async and enable get_tensor_async
This allows some additional CPU/GPU overlap for large pp workloads. Also seems
to help a bit for token gen, maybe getting rid of a small bubble between
graph_compute and get_tensor.
Async set and copy functions seem to be very rarely used, so I didn't enable
them because I didn't have a good way to test them.
The async commands need to be ordered against each other, so put them all on
the compute queue. The non-async commands still use the transfer queue.
The fence for graph_compute/get_tensor_async is submitted and waited on in
ggml_vk_synchronize.
* fix thread safety errors
* teardown context cleanly
* Handle async read to non-pinned dst
This change combines the rms_norm+mul and rope+view+set_rows fusions to
allow fusing the whole sequence together. This comes up in Qwen3, Bailing,
and some other models.
The std::map pipeline_flash_attn_f32_f16 could be searched and inserted at the
same time, which needs to hold the lock. To be safe, hold the lock for all of
ggml_vk_load_shaders.
* vulkan: remove the need for the dryrun
Allocate pipelines and descriptor sets when requested.
Reallocate the prealloc buffers when needed, and flush any pending work
before reallocating.
For rms_partials and total_mul_mat_bytes, use the sizes computed the last time
the graph was executed.
* remove dryrun parameters
* vulkan: fuse mul_mat+add and mul_mat_id+add_id
The fusion is only applied for the mat-vec mul paths.
* Apply suggestions from code review
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* fix 32b build
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This pattern appears in a lot of models, the rope operation is applied right
before storing into the KV cache (usually on the K tensor).
Add a path to some of the rope shaders that computes the destination address
based on the set_rows tensor. Compile variants of the shader with D_TYPE of
f16 (the usual KV cache type).
Add a src3 operand to ggml_vk_op_f32 - sometimes rope uses three srcs and needs
the fourth for the row indices.
Add fused_ops_write_mask to indicate which intermediate tensors need to write
their results to memory. Skipping writing the roped K value helps to allow more
nodes to run concurrently.
Add logic to ggml_vk_graph_optimize to make ROPE+VIEW+SET_ROWS consecutive. It
rarely starts out that way in the graph.
Add new backend tests.
