@@ -33,6 +33,7 @@ def test_prompts_aligned_with_tkv_boundaries(model: str, backend: str,
3333 steps_add_reqs = [0 , 0 , 0 ] # add all requests in the beginning
3434 available_blocks = - 1 # no restriction
3535 max_num_seqs = 2
36+ max_model_len = 256
3637
3738 checked_steps = [
3839 {
@@ -170,6 +171,7 @@ def test_prompts_aligned_with_tkv_boundaries(model: str, backend: str,
170171 steps_add_reqs = steps_add_reqs ,
171172 checked_steps = checked_steps ,
172173 max_num_seqs = max_num_seqs ,
174+ max_model_len = max_model_len ,
173175 available_blocks = available_blocks ,
174176 use_cb = True ,
175177 )
@@ -197,6 +199,7 @@ def test_prompts_misaligned_with_tkv_boundaries(
197199 steps_add_reqs = [0 , 0 , 0 ] # add all requests in the beginning
198200 available_blocks = - 1 # no restriction
199201 max_num_seqs = 2
202+ max_model_len = 256
200203
201204 checked_steps = [
202205 {
@@ -332,6 +335,7 @@ def test_prompts_misaligned_with_tkv_boundaries(
332335 steps_add_reqs = steps_add_reqs ,
333336 checked_steps = checked_steps ,
334337 max_num_seqs = max_num_seqs ,
338+ max_model_len = max_model_len ,
335339 available_blocks = available_blocks ,
336340 use_cb = True ,
337341 )
@@ -358,6 +362,7 @@ def test_two_sequences_finish_same_time_as_new_arrive(
358362 steps_add_reqs = [0 , 0 , 31 ]
359363 available_blocks = - 1 # no restriction
360364 max_num_seqs = 2
365+ max_model_len = 256
361366
362367 checked_steps = [
363368 {
@@ -470,6 +475,7 @@ def test_two_sequences_finish_same_time_as_new_arrive(
470475 steps_add_reqs = steps_add_reqs ,
471476 checked_steps = checked_steps ,
472477 max_num_seqs = max_num_seqs ,
478+ max_model_len = max_model_len ,
473479 available_blocks = available_blocks ,
474480 use_cb = True ,
475481 )
@@ -496,6 +502,7 @@ def test_new_sequence_joins_during_decode(model: str, backend: str,
496502 steps_add_reqs = [0 , 0 , 32 , 131 ]
497503 available_blocks = - 1 # no restriction
498504 max_num_seqs = 4
505+ max_model_len = 256
499506
500507 checked_steps = [
501508 {
@@ -707,6 +714,7 @@ def test_new_sequence_joins_during_decode(model: str, backend: str,
707714 steps_add_reqs = steps_add_reqs ,
708715 checked_steps = checked_steps ,
709716 max_num_seqs = max_num_seqs ,
717+ max_model_len = max_model_len ,
710718 available_blocks = available_blocks ,
711719 use_cb = True ,
712720 )
@@ -731,6 +739,7 @@ def test_prompt_too_long_for_current_tkv(model: str, backend: str,
731739 steps_add_reqs = [0 , 0 ]
732740 available_blocks = - 1 # no restriction
733741 max_num_seqs = 2
742+ max_model_len = 256
734743
735744 checked_steps = [
736745 {
@@ -854,6 +863,7 @@ def test_prompt_too_long_for_current_tkv(model: str, backend: str,
854863 steps_add_reqs = steps_add_reqs ,
855864 checked_steps = checked_steps ,
856865 max_num_seqs = max_num_seqs ,
866+ max_model_len = max_model_len ,
857867 available_blocks = available_blocks ,
858868 use_cb = True ,
859869 )
@@ -879,6 +889,7 @@ def test_requested_tokens_not_fitting_remaining_space(
879889 steps_add_reqs = [0 , 0 , 0 ]
880890 available_blocks = - 1 # no restriction
881891 max_num_seqs = 2
892+ max_model_len = 256
882893
883894 checked_steps = [
884895 {
@@ -1039,6 +1050,7 @@ def test_requested_tokens_not_fitting_remaining_space(
10391050 steps_add_reqs = steps_add_reqs ,
10401051 checked_steps = checked_steps ,
10411052 max_num_seqs = max_num_seqs ,
1053+ max_model_len = max_model_len ,
10421054 available_blocks = available_blocks ,
10431055 use_cb = True ,
10441056 )
@@ -1067,6 +1079,8 @@ def test_requests_use_all_available_blocks(model: str, backend: str,
10671079 # total number of blocks needed if scheduled together : 4 * (1 + 1) = 8
10681080 available_blocks = 8
10691081 max_num_seqs = 4
1082+ max_model_len = 256
1083+
10701084 checked_steps = [
10711085 {
10721086 "step" : 0 ,
@@ -1170,6 +1184,7 @@ def test_requests_use_all_available_blocks(model: str, backend: str,
11701184 steps_add_reqs = steps_add_reqs ,
11711185 checked_steps = checked_steps ,
11721186 max_num_seqs = max_num_seqs ,
1187+ max_model_len = max_model_len ,
11731188 available_blocks = available_blocks ,
11741189 use_cb = True ,
11751190 )
@@ -1199,6 +1214,8 @@ def test_requests_use_more_than_available_blocks(
11991214 # total number of blocks needed if scheduled together : 4 * (1 + 1) = 8
12001215 available_blocks = 4
12011216 max_num_seqs = 4
1217+ max_model_len = 256
1218+
12021219 checked_steps = [
12031220 {
12041221 "step" : 0 ,
@@ -1327,6 +1344,7 @@ def test_requests_use_more_than_available_blocks(
13271344 steps_add_reqs = steps_add_reqs ,
13281345 checked_steps = checked_steps ,
13291346 max_num_seqs = max_num_seqs ,
1347+ max_model_len = max_model_len ,
13301348 available_blocks = available_blocks ,
13311349 use_cb = True ,
13321350 )
0 commit comments