CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-28 08:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: use FA + max. GPU layers by default (#15434)

Browse Source 
* llama: use max. GPU layers by default, auto -fa

* ggml-backend: abort instead of segfault
This commit is contained in:
Johannes Gäßler
2025-08-30 16:32:10 +02:00
committed by GitHub
parent 38ad381f9f
commit e81b8e4b7f
19 changed files with 235 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
tools/server/tests/unit/test_ctx_shift.py
View File

@@ -15,25 +15,26 @@ Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deseru
def create_server():
global server
server = ServerPreset.tinyllama2()
server.n_ctx = 256
server.n_ctx = 512
server.n_slots = 2
server.n_predict = 128
def test_ctx_shift_enabled():
# the prompt is 301 tokens
# the slot context is 256/2 = 128 tokens
# the prompt is truncated to keep the last 109 tokens
# 64 tokens are generated thanks to shifting the context when it gets full
# the slot context is 512/2 = 256 tokens
# the prompt is truncated to keep the last (301 - 256/2) = 173 tokens
# 96 tokens are generated thanks to shifting the context when it gets full
global server
server.enable_ctx_shift = True
server.start()
res = server.make_request("POST", "/completion", data={
"n_predict": 64,
"n_predict": 96,
"prompt": LONG_TEXT,
})
assert res.status_code == 200
assert res.body["timings"]["prompt_n"] == 109
assert res.body["timings"]["predicted_n"] == 64
assert res.body["timings"]["prompt_n"] == 173
assert res.body["timings"]["predicted_n"] == 96
assert res.body["truncated"] is True