Stats/Logs (#2165)

* rename

* stats

* stats

* update state

* update state

Author: Qubitium

gptqmodel/looper/gptq_processor.py

@@ -185,6 +185,9 @@ def process(
 
         wq, q_scales, q_zeros, q_g_idx, duration, avg_loss, damp_percent, nsamples = g.quantize()
 
+        workspace_summary = getattr(g, "_borrow_workspace_last_summary", None)
+        workspace_totals = getattr(g, "_borrow_workspace_totals", None)
+
         module.stream_state_payload_to_cpu(
             {
                 "q_scales": q_scales,
@@ -235,6 +238,25 @@ def process(
             PROCESS_USED_MEMORY: self.device_memory_report(),
         }
 
+        if workspace_summary:
+            requests = int(workspace_summary.get("requests", 0) or 0)
+            if requests:
+                hit_rate = float(workspace_summary.get("hit_rate", 0.0) or 0.0)
+                chunk_rows = workspace_summary.get("chunk_rows")
+                stat["workspace_cache_requests"] = str(requests)
+                stat["workspace_cache_hit_rate"] = f"{hit_rate:.1%}"
+                stat["workspace_stage_dtype"] = workspace_summary.get("staging_dtype", "")
+                if chunk_rows is not None:
+                    stat["workspace_chunk_rows"] = str(chunk_rows)
+        if workspace_totals:
+            total_requests = int(workspace_totals.get("requests", 0) or 0)
+            if total_requests:
+                cumulative_hit_rate = (
+                    float(workspace_totals.get("materialized_hits", 0) or 0.0) / total_requests
+                )
+                stat["workspace_total_requests"] = str(total_requests)
+                stat["workspace_total_hit_rate"] = f"{cumulative_hit_rate:.1%}"
+
         if self.qcfg.dynamic is not None:
             stat["dynamic"] = self.qcfg.dynamic_get(layer_name=module.full_name)
 
@@ -244,6 +266,8 @@ def process(
         # Log the new row
         self.log_new_row(stat)
 
+        g.log_workspace_stats(context="gptq_process")
+
         if self.calculate_w_wq_diff:
             # diff in float32
             w_wq_diff = module.weight.data.to(dtype=torch.float32) - wq.to(dtype=torch.float32)

gptqmodel/looper/loop_processor.py

@@ -439,9 +439,11 @@ def device_memory_report(self) -> str:
             return "n/a"
 
         def _format_gib(value: float) -> str:
-            text = f"{value:.1f}"
-            if text.endswith(".0"):
+            text = f"{value:.2f}"
+            if text.endswith("00"):
                 text = text[:-2]
+            elif text.endswith("0"):
+                text = text[:-1]
             return f"{text}G"
 
         grouped: Dict[str, List[Tuple[str, float, int]]] = {}