Merge branch 'main' of https://github.com/rwilliamspbg-ops/Mohawk-Inference-Engine

Author: Sovereign Map Test Suite

prototype/test_correctness_suite.py

@@ -12,6 +12,7 @@
 """
 
 import sys
+import time
 import numpy as np
 import pytest
 from prototype.model_tools import ToyModel
@@ -319,19 +320,28 @@ def test_throughput_consistency(self):
         for bs in batch_sizes:
             x = np.random.randn(bs, 8).astype(np.float32)
 
-            # Measure time for single-node
-            start = np.datetime64('now')
-            _ = single_node_forward(model, x[0])  # First element only
-            end = np.datetime64('now')
+            # Warm up to avoid first-run JIT/cache effects
+            for _ in range(3):
+                _ = single_node_forward(model, x[0])
 
-            latency = (end - start).astype(np.float64) * 1e9  # nanoseconds
+            # Take best-of-N to reduce OS scheduling noise
+            samples = []
+            for _ in range(10):
+                start = time.perf_counter_ns()
+                _ = single_node_forward(model, x[0])  # First element only
+                end = time.perf_counter_ns()
+                samples.append(end - start)
 
-            latencies.append(latency)
+            best_latency = float(min(samples))  # nanoseconds
+            latencies.append(best_latency)
 
         # Latencies should be within 20% of each other (allowing for variance)
         min_latency = min(latencies)
         max_latency = max(latencies)
 
+        if min_latency < 100:
+            pytest.skip("Timer resolution too low to measure latency reliably")
+        
         assert (max_latency - min_latency) / min_latency < 0.2, \
             f"Latency variance too high: {min_latency} vs {max_latency}"
 

requirements.txt

@@ -0,0 +1,9 @@
+transformers>=4.30.0
+accelerate>=0.20.0
+bitsandbytes>=0.41.0
+optimum>=1.6.0
+vllm>=0.4.0
+kubernetes>=27.0.0
+cryptography>=41.0.0
+pyoqs>=0.9.0  # For PQC liboqs binding
+prometheus-client>=0.15.0