transformers
Fix Glm4MoeIntegrationTest: split into 3 tests, switch to GLM-4.5-Air
#48820
Merged
Merged
Fix Glm4MoeIntegrationTest: split into 3 tests, switch to GLM-4.5-Air #48820
ydshieh
merged
19 commits into
main
from
fix-glm4-moe-integration-test
Fix Glm4MoeIntegrationTest: device_map=auto, bfloat16, MemoryCleanupM…
82cd4e0a
Switch to 4-bit quantization: bfloat16 (716 GiB) OOMs both runners
45e742d9
Fix Glm4MoeIntegrationTest: offload_folder + remove 4bit quantization
4bcf0cc1
Comment out generate calls in Glm4MoeIntegrationTest, keep only loading
29b77cb9
[Temp] Fail check_bad_commit early to skip find-bad-commit job
9e09b172
Uncomment dynamic generate in Glm4MoeIntegrationTest
01f33d87
Switch to GLM-4.5-Air (106B/12B active) and comment out generate for …
ae1c0801
Split Glm4MoeIntegrationTest into 3 test methods (dynamic/static/comp…
cbba48a4
Reduce to 5 tokens, enable asserts (EXPECTED_TEXT_COMPLETION=None to …
ad088688
Fix Glm4MoeIntegrationTest: GLM-4.5-Air, 3 separate tests, expected v…
6c796a1b
Move torch import inside is_torch_available() guard
2501c212
Use dtype=torch.bfloat16 explicitly instead of dtype="auto"
7da440b1
Restore 70% per-GPU max_memory cap to avoid CUDA OOM on multi-GPU run…
8c2a5de1
Add CPU memory cap via get_cpu_ram_total_gib to avoid disk offload on…
75cb3670
Use 90% of CPU RAM to avoid runner exit code 137
49d53097
Fix ruff formatting in glm4_moe max_memory cpu line
6519ba3c
Use dtype="auto" to preserve native mixed precision from checkpoint
a27f0d77
Revert dummy check_bad_commit.py change
9496a2e1
Remove redundant @slow/@require_torch_accelerator from test methods
738258b1
ydshieh
changed the title
Login to write a write a comment.
Login via GitHub