Upload FINDINGS.md with huggingface_hub

FINDINGS.md

@@ -48,9 +48,19 @@ by measurement:
 
 What stands is empirical, not theoretical: **no rebalancing of CPU against GPU
 within 92 GB of VRAM improves anything**, across placements spanning 60–88 GB of
-VRAM and 4–33 GB of host traffic per token. The practical conclusion holds —
-6–8 tok/s needs more VRAM (4× A40, ~$1.80/h), not more tuning — but the reason
-it holds is not established, and should not be presented as if it were.
+VRAM and 4–33 GB of host traffic per token.
+
+**The GPUs are not decorative, though — verified.** Running with `-ngl 0`, so
+that essentially nothing sits in VRAM (1265 MiB and 269 MiB), gives **0.65
+tok/s** against 2.51 for the same prompt with autofit. The two A40s are worth
+**3.9×**. That check was run specifically because, if the GPUs had contributed
+nothing, "buy more VRAM" would have been the wrong advice.
+
+So the direction — fit more of the model in VRAM — is sound. The magnitude is
+not predictable from here: a naive linear fit through (0 GB → 0.65) and
+(92 GB → 2.94) lands near 5.4 tok/s at full residency, but removing the host
+path entirely should be super-linear, since it swaps a mixed regime for pure
+VRAM at 696 GB/s. The honest answer is "substantially better, number unknown".
 
 ## What did NOT work, and why