feat(mm): log activation multiplier in DENOISE_MEM for self-contained calibration

Each DENOISE_MEM record now carries the 'mult' that produced its estimate, so the implied per-arch multiplier can be back-solved from a single line without external state. This matters during iterative calibration, where the multiplier changes between runs that may share one log file.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Your Name

invokeai/backend/util/denoise_working_memory.py

@@ -242,6 +242,7 @@ def end_denoise_measure(
                     "px_w": int(pixel_width),
                     "batch": int(batch_size),
                     "elt": int(element_size),
+                    "mult": ACTIVATION_MULTIPLIER["unet"] if is_unet else ACTIVATION_MULTIPLIER["dit"],
                     "estimate_mb": round(estimate_bytes / MB, 1),
                     "measured_peak_mb": round(measured / MB, 1),
                     "resident_before_mb": round(token / MB, 1),

tests/backend/util/test_denoise_working_memory.py

@@ -1,5 +1,6 @@
 """Tests for the architecture-scaled denoise working-memory estimator."""
 
+import json
 import types
 
 import torch
@@ -189,6 +190,11 @@ def debug(self, msg: str) -> None:
     if torch.cuda.is_available():
         assert token is not None
         assert len(log.records) == 1 and log.records[0].startswith("DENOISE_MEM ")
+        payload = json.loads(log.records[0][len("DENOISE_MEM ") :])
+        # The record must be self-describing for calibration: it carries the multiplier that produced
+        # the estimate (flux2 -> "dit" family), so back-solving M needs no external state.
+        assert payload["mult"] == ACTIVATION_MULTIPLIER["dit"]
+        assert payload["label"] == "flux2" and "measured_peak_mb" in payload
     else:
         assert token is None
         assert log.records == []