Age | Commit message (Collapse) | Author |
|
AVX512 (#1119)
|
|
|
|
|
|
|
|
* Unit test for quantization functions
Use the ggml_internal_get_quantize_fn function to loop through all
quantization formats and run a sanity check on the result.
Also add a microbenchmark that times these functions directly without
running the rest of the GGML graph.
* test-quantize-fns: CI fixes
Fix issues uncovered in CI
- need to use sizes divisible by 32*8 for loop unrolling
- use intrinsic header that should work on Mac
* test-quantize: remove
Per PR comment, subsumed by test-quantize-fns
* test-quantize: fix for q8_0 intermediates
|
|
* print timings on ctrl+c exit
* remove redundant free memory call.
* add global pointer to ctx.
|
|
* set default n_batch to 512 when using BLAS
* spacing
* alternate implementation of setting different n_batch for BLAS
* set n_batch to 512 for all cases
|
|
* Fix build under Windows when enable BUILD_SHARED_LIBS
* Make AVX512 test on Windows to build the shared libs
|
|
|
|
* ggml : prefer vzip to vuzp
This way we always use the same type of instruction across all quantizations
* ggml : alternative Q4_3 implementation using modified Q8_0
* ggml : fix Q4_3 scalar imlpementation
* ggml : slight improvement of Q4_3 - no need for loop unrolling
* ggml : fix AVX paths for Q8_0 quantization
|
|
* AVX2 optimization for vec_dot_q4_3_q8_0 and refactoring
* finish AVX vectorization of quantize_row_q8_0
* Rename hsum_int_8 to hsum_i32_8
|
|
Experience (#1107)
* Moving parameters to separate lines for readability.
* Increasing repeate_penalty to 1.1 to make alpaca more usable by default.
* Adding trailing newline.
|
|
embedding and kv_cache (#1105)
* reserve correct size for logits
* add functions to get and set the whole llama state:
including rng, logits, embedding and kv_cache
* remove unused variables
* remove trailing whitespace
* fix comment
|
|
* Improve cuBLAS performance by using a memory pool
* Move cuda specific definitions to ggml-cuda.h/cu
* Add CXX flags to nvcc
* Change memory pool synchronization mechanism to a spin lock
General code cleanup
|
|
|
|
* fix: ld link test-tokenizer-0 error
```
cmake3 --build . --config Release
[ 5%] Built target ggml
[ 16%] Built target llama
[ 22%] Linking CXX executable ../bin/test-tokenizer-0
../libllama.a(ggml.c.o):在函数‘ggml_graph_compute’中:
ggml.c:(.text+0xf2db):对‘pthread_create’未定义的引用
ggml.c:(.text+0xf9d4):对‘pthread_join’未定义的引用
collect2: error: ld returned 1 exit status
gmake[2]: *** [bin/test-tokenizer-0] 错误 1
gmake[1]: *** [tests/CMakeFiles/test-tokenizer-0.dir/all] 错误 2
gmake: *** [all] 错误 2
```
* Update CMakeLists.txt
* Update CMakeLists.txt
* Update CMakeLists.txt
|
|
* examples : evaluate tokens in batches after swapping context
* Update examples/main/main.cpp
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
|
|
because their value is stored in buf and overwritten by memcpy
|
|
* A faster version for Q4_1 x Q8_0 dot products
The idea nehind being that Q8_0 quantized
values get used many times in the matrix multiplications
where they are involved. In the current implementations,
when we are evaluating the dot products, we need to compute
the sum of the quants in the Q8_0 vector, so the same
operation is repeated many times. Here we pre-compute
the sum during Q8_0 quantization, store it in the
now modified block_q8_0 struct, and then reuse this
result in the subsequent dot products.
In a synthetic benchmark (just compute a bunch of dot
products), this change speeds up the Q4_1 * Q8_0 dot
product by 80%, making the performance identical to
Q4_0 * Q8_0.
In practical application, I see a ~15% gain in speed for
token prediction on M2, and ~5% gain on Ryzen 7950X.
The speed gain in the prompt evaluation is much bigger
(around 50%).
I have only done the change for the scalar version,
ARM_NEON, and AVX2, so we still need an AVX implementation.
* Cleaning up
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
|
|
|
|
|
|
* Add ggml-model-*.bin checksums for 7B, 13B, 30B
* Add ggml-model-*.bin checksums for 65B
---------
Co-authored-by: Pavol Rusnak <pavol@rusnak.io>
|
|
|
|
|
|
|
|
|
|
Broke it during conflict resolution in last PR
|
|
* Multi-threading quantization.
Not much gain for simple quantizations, bit it will be important
for quantizations that require more CPU cycles.
* Multi-threading for quantize-stats
It now does the job in ~14 seconds on my Mac for
Q4_0, Q4_1 and Q4_2. Single-threaded it was taking
more than 2 minutes after adding the more elaborate
version of Q4_2.
* Reviewer comments
* Avoiding compiler confusion
After changing chunk_size to const int as suggested by
@ggerganov, clang and GCC starting to warn me that I don't
need to capture it in the lambda. So, I removed it from the
capture list. But that makes the MSVC build fail. So,
making it a constexpr to make every compiler happy.
* Still fighting with lambda captures in MSVC
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
|
|
|
|
CI (#1074)
[Accelerate](https://developer.apple.com/documentation/accelerate) is an Apple framework which can only be used on macOS, and the CMake build [ignores](https://github.com/ggerganov/llama.cpp/blob/master/CMakeLists.txt#L102) the `LLAMA_ACCELERATE` variable when run on non-Apple platforms. This implies setting `LLAMA_ACCELERATE` is a no-op on Ubuntu and can be removed.
This will reduce visual noise in CI check results (in addition to reducing the number of checks we have to run for every PR). Right now every sanitized build is duplicated twice for no good reason (e.g., we have `CI / ubuntu-latest-cmake-sanitizer (ADDRESS, Debug, ON)` and `CI / ubuntu-latest-cmake-sanitizer (ADDRESS, Debug, OFF)`).
|
|
|
|
|
|
|
|
|
|
* Q4_2 quantization with rmse-optimized scale and quants
For quantize-stats we get
q4_2: rmse 0.00159301, maxerr 0.17480469, 95pct<0.0030, median<0.0012
For 7B perplexity with BLAS enabled we get 6.2038 after 655 chunks.
Quantization is slow (~90 seconds on my Mac for 7B) as not
multi-threaded as in PR #896.
* ggml : satisfy the sanitizer builds
Not sure why this makes them fail
* Better follow ggml conventions for function names
* Fixed type as per reviewer comment
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
|
|
* ggml : use 8-bit precision for Q4_1 intermediate results (ARM)
* ggml : optimize ggml_vec_dot_q4_1_q8_0() via vmalq_n_f32
56 ms/token with Q4_1 !
* ggml : AVX2 implementation of ggml_vec_dot_q4_1_q8_0 (#1051)
* gitignore : ignore ppl-*.txt files
---------
Co-authored-by: slaren <2141330+slaren@users.noreply.github.com>
|
|
|
|
* Q4 cleanup
* Remove unused AVX512 Q4_0 code
|
|
|
|
* Multi-threaded ggml_cpy
* Update ggml.c
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Also fix wdata offset in ggml_compute_forward_add_q_f32
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
|
|
* ggml : Q4_2 ARM
* ggml : add ggml_is_quantized()
* llama : update llama_type_name() with Q4_2 entry
* ggml : speed-up q4_2
- 4 threads: ~100ms -> ~90ms
- 8 threads: ~55ms -> ~50ms
* ggml : optimize q4_2 using vmlaq_n_f32 + vmulq_n_f32
|
|
Had a background process that was messing with the timings
|
|
|
|
|
|
On my Mac, the direct Q4_1 product is marginally slower
(~69 vs ~55 us for Q4_0). The SIMD-ified ggml version
is now almost 2X slower (~121 us).
On a Ryzen 7950X CPU, the direct product for Q4_1 quantization
is faster than the AVX2 implementation (~60 vs ~62 us).
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
|
|
|
|
|
|
|
|
|
|
this came up when trying to convert the gpt4all-lora-unfiltered-quantized.bin file
|