tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

chore: adjust log

Browse Source
This commit is contained in:
NanoCode012
2026-02-25 17:49:57 +07:00
parent 4b2f568ee0
commit ca822cd24c
1 changed files with 12 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/axolotl/monkeypatch/moe_quant.py
View File

@@ -81,7 +81,13 @@ def patch_moe_quantization_on_load(cfg):
def _patched_set_param_for_module(model, target_name, param_value, *args, **kwargs): def _patched_set_param_for_module(model, target_name, param_value, *args, **kwargs):
if _first_call[0]: if _first_call[0]:
LOG.info("MoE quant patch: set_param_for_module intercepted (first call)") LOG.info(
"MoE quant patch: set_param_for_module intercepted (first call) "
"(alloc=%.2f GiB, reserved=%.2f GiB, max_alloc=%.2f GiB)",
torch.cuda.memory_allocated() / 1024**3,
torch.cuda.memory_reserved() / 1024**3,
torch.cuda.max_memory_allocated() / 1024**3,
)
_first_call[0] = False _first_call[0] = False
original_set_param(model, target_name, param_value, *args, **kwargs) original_set_param(model, target_name, param_value, *args, **kwargs)
@@ -116,6 +122,11 @@ def patch_moe_quantization_on_load(cfg):
transformers.core_model_loading.set_param_for_module = _patched_set_param_for_module transformers.core_model_loading.set_param_for_module = _patched_set_param_for_module
_moe_load_state["patched"] = True _moe_load_state["patched"] = True
LOG.info(
"Pre-load GPU memory: alloc=%.2f GiB, reserved=%.2f GiB",
torch.cuda.memory_allocated() / 1024**3,
torch.cuda.memory_reserved() / 1024**3,
)
LOG.info( LOG.info(
"Activated MoE loading-time quantization patch " "Activated MoE loading-time quantization patch "
"(quant_type=%s, compress_statistics=%s)", "(quant_type=%s, compress_statistics=%s)",