diff options
author | Kawrakow <48489457+ikawrakow@users.noreply.github.com> | 2023-04-20 19:42:27 +0200 |
---|---|---|
committer | GitHub <noreply@github.com> | 2023-04-20 20:42:27 +0300 |
commit | 38de86a7114c97ecf3644e3a60159f1ed893e1b0 (patch) | |
tree | fc6b90dd99825ce4e745304aab484b85903949c0 /ggml.c | |
parent | e0305ead3a072db9c08b35c9600c49273b38a4b5 (diff) |
llama : multi-threaded quantization (#1075)
* 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>
Diffstat (limited to 'ggml.c')
-rw-r--r-- | ggml.c | 27 |
1 files changed, 27 insertions, 0 deletions
@@ -12189,6 +12189,33 @@ size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * return (n/QK4_3*sizeof(block_q4_3)); } +size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) { + size_t result = 0; + switch (type) { + case GGML_TYPE_Q4_0: + { + GGML_ASSERT(start % QK4_0 == 0); + block_q4_0 * block = (block_q4_0*)dst + start / QK4_0; + result = ggml_quantize_q4_0(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_1: + { + GGML_ASSERT(start % QK4_1 == 0); + block_q4_1 * block = (block_q4_1*)dst + start / QK4_1; + result = ggml_quantize_q4_1(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_2: + { + GGML_ASSERT(start % QK4_2 == 0); + block_q4_2 * block = (block_q4_2*)dst + start / QK4_2; + result = ggml_quantize_q4_2(src + start, block, n, n, hist); + } break; + default: + assert(false); + } + return result; +} + //////////////////////////////////////////////////////////////////////////////// int ggml_cpu_has_avx(void) { |