* metal: better memory alloc w/ concurrency dispatch
The ggml-alloc should only free tensors at memory barriers.
* ggml-alloc: avoid return silently
In certain cases, the allocate_node() function may silently return
without performing any memory allocation.
The GGML memory allocator consistently places a tensor within the
optimal-fit memory block, which is the smallest block capable of
accommodating the tensor's size. During the measurement phase, the final
block is generously sized, ensuring it never qualifies as the
optimal-fit block as long as there exists another block capable of
accommodating the tensor. Nevertheless, in the evaluation phase, the
last block is constrained in size and could potentially qualify as the
optimal-fit block. Consequently, there exists the possibility of a
tensor being allocated to a different region during evaluation, leading
to more memory fragmentation in our scratch buffer.
This recent commit guarantees uniform behavior of the allocator across
both the measurement and evaluation phases, eliminating discrepancies
between the two.
* metal: matrix-matrix multiplication kernel
This commit removes MPS and uses custom matrix-matrix multiplication
kernels for all quantization types. This commit also adds grouped-query
attention to support llama2 70B.
* metal: fix performance degradation from gqa
Integers are slow on the GPU, and 64-bit divides are extremely slow.
In the context of GQA, we introduce a 64-bit divide that cannot be
optimized out by the compiler, which results in a decrease of ~8% in
inference performance. This commit fixes that issue by calculating a
part of the offset with a 32-bit divide. Naturally, this limits the
size of a single matrix to ~4GB. However, this limitation should
suffice for the near future.
* metal: fix bugs for GQA and perplexity test.
I mixed up ne02 and nb02 in previous commit.
* metal: concurrently dispatch commands
Function `ggml_metal_graph_find_concurrency` will run and write
commands that can be issued concurrently to metal context `concur_list`
array, when `ggml_metal_graph_compute` is called for the first time.
* metal: don't call find_concurrency automatically.
* metal : code style changes
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* metal: use uint16_t instead of uint8_t.
Apple GPU doesn't like uint8_t. For every operation on uint8_t
the gpu need to copy the uint8_t to an empty 16 bit register, then
it can issue other instructions.
For the matrix-vector multiplication kernel only, we observed a
340~350 GB/s memory read speed on M1 Max after this commit, which is
very close to the reported hardware limit.
* metal: update rms_norm kernel
This commit double the speed of rms_norm operations by using 512 threads
per threadgroup, combining with SIMD primitives to minimize the need for
thread group barriers.
* metal: use template to reduce size
Revert modifications on block_q4_0 and block_q4_1.
