Skip to content

Commit b0023dc

Browse files
committed
tests still unreliable with Ollama version in GitHub CI
These tests should work and do work locally. But they fail in GitHub CI – for an unknown reason that almost certainly is in Ollama, not in our code.
1 parent 08160b7 commit b0023dc

File tree

1 file changed

+3
-3
lines changed

1 file changed

+3
-3
lines changed

tests/tollamaChat.m

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -50,7 +50,7 @@ function extremeTopK(testCase)
5050
%% This should work, and it does on some computers. On others, Ollama
5151
%% receives the parameter, but either Ollama or llama.cpp fails to
5252
%% honor it correctly.
53-
% testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
53+
testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
5454

5555
% setting top-k to k=1 leaves no random choice,
5656
% so we expect to get a fixed response.
@@ -65,7 +65,7 @@ function extremeMinP(testCase)
6565
%% This should work, and it does on some computers. On others, Ollama
6666
%% receives the parameter, but either Ollama or llama.cpp fails to
6767
%% honor it correctly.
68-
% testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
68+
testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
6969

7070
% setting min-p to p=1 means only tokens with the same logit as
7171
% the most likely one can be chosen, which will almost certainly
@@ -81,7 +81,7 @@ function extremeTfsZ(testCase)
8181
%% This should work, and it does on some computers. On others, Ollama
8282
%% receives the parameter, but either Ollama or llama.cpp fails to
8383
%% honor it correctly.
84-
% testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
84+
testCase.assumeTrue(false,"disabled due to Ollama/llama.cpp not honoring parameter reliably");
8585

8686
% setting tfs_z to z=0 leaves no random choice, but degrades to
8787
% greedy sampling, so we expect to get a fixed response.

0 commit comments

Comments
 (0)