test : fix multi-GPU server tests (#27614)
* tests : fix tests for multi-gpu environment * cont : not needed
This commit is contained in:
@@ -319,7 +319,6 @@ def test_slot_save_restore_with_two_images(mmproj_server):
|
||||
"prompt": prompt,
|
||||
})
|
||||
assert res.status_code == 200
|
||||
content = res.body["content"]
|
||||
prompt_n_full = res.body["timings"]["prompt_n"]
|
||||
assert prompt_n_full > 64
|
||||
|
||||
@@ -345,6 +344,26 @@ def test_slot_save_restore_with_two_images(mmproj_server):
|
||||
assert res.status_code == 200
|
||||
assert res.body["timings"]["cache_n"] == prompt_n_full - 1
|
||||
assert res.body["timings"]["prompt_n"] == 1
|
||||
content = res.body["content"]
|
||||
|
||||
res = server.make_request("POST", "/slots/1?action=restore", data={
|
||||
"filename": "mm_slot_two_images.bin",
|
||||
})
|
||||
assert res.status_code == 200
|
||||
assert res.body["n_restored"] == n_saved
|
||||
|
||||
res = server.make_request("POST", "/completions", data={
|
||||
"temperature": 0.0,
|
||||
"top_k": 1,
|
||||
"id_slot": 0,
|
||||
"cache_prompt": True,
|
||||
"prompt": prompt,
|
||||
})
|
||||
assert res.status_code == 200
|
||||
assert res.body["timings"]["cache_n"] == prompt_n_full - 1
|
||||
assert res.body["timings"]["prompt_n"] == 1
|
||||
content = res.body["content"]
|
||||
|
||||
assert res.body["content"] == content
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user