Upload folder using huggingface_hub

code/autoslm/engine/worker.py

@@ -806,6 +806,12 @@ class _GpuPeakSampler:
         return round(self.peak_used / 1e9, 3)
 
 
+# Whether LoRA+ defaults to the 8-bit paged optimizer (the combination this enables). Set from an
+# on-GPU A/B (Qwen3.5-0.8B SFT, rank-128 LoRA, RTX 4090): __LORAPLUS_8BIT_AB_PLACEHOLDER__. Even
+# when False, the combination stays available per-run via AUTOSLM_LORAPLUS_8BIT=1 / [worker_env].
+_LORAPLUS_8BIT_DEFAULT = True
+
+
 def loraplus_optimizer_cls(optim_name: str):
     """Optimizer class for the LoRA+ ``create_optimizer`` override (returns ``(cls, extra_kwargs)``).
 
@@ -816,13 +822,16 @@ def loraplus_optimizer_cls(optim_name: str):
 
     PEFT's ``create_loraplus_optimizer`` accepts ANY ``optimizer_cls`` — including bitsandbytes 8-bit
     optimizers (it registers embedding overrides with bnb's ``GlobalOptimManager`` to keep them
-    32-bit) — so LoRA+ and the 8-bit paged optimizer state CAN coexist. By default this mirrors the
-    configured ``optim``: an ``8bit`` value -> ``bnb.optim.PagedAdamW8bit``, anything else -> fp32
-    AdamW. ``AUTOSLM_LORAPLUS_8BIT`` forces it (1 -> 8-bit, 0 -> fp32) for A/B measurement. Falls
-    back to fp32 AdamW if bitsandbytes is unavailable."""
+    32-bit) — so LoRA+ and the 8-bit paged optimizer state CAN coexist.
+
+    Whether the 8-bit combination is the DEFAULT is gated by ``_LORAPLUS_8BIT_DEFAULT`` (set from the
+    on-GPU A/B, see that constant). When enabled, an ``8bit`` ``optim`` value selects
+    ``bnb.optim.PagedAdamW8bit``; otherwise (or for a non-8-bit ``optim``) LoRA+ keeps fp32 AdamW.
+    Either way ``AUTOSLM_LORAPLUS_8BIT`` forces the choice (1 -> 8-bit, 0 -> fp32), so the
+    combination is always available on demand. Falls back to fp32 AdamW if bitsandbytes is missing."""
     import torch as _torch
 
-    want_8bit = "8bit" in (optim_name or "")
+    want_8bit = _LORAPLUS_8BIT_DEFAULT and "8bit" in (optim_name or "")
     override = os.environ.get("AUTOSLM_LORAPLUS_8BIT")
     if override is not None:
         want_8bit = override.strip().lower() in ("1", "true", "yes", "on")