Calibrate MTP layer with runtime-aware AWQ

Replace the MTP shard with runtime-aware AWQ calibration from SWE-chat code-generation activations. The 24-session SGLang evaluation improved aggregate throughput by 3.97%.

CHANGELOG.md

@@ -0,0 +1,13 @@
+# Changelog
+
+## 2026-07-21
+
+Replaced `model-00041-of-00041.safetensors` with a runtime-aware AWQ-calibrated MTP layer. Calibration used code-generation activations from session-disjoint SWE-chat data and selected scales with the deployed SGLang CUTLASS and DeepGEMM kernels.
+
+On 24 untouched SWE-chat sessions at concurrency 8 with three speculative steps and four draft tokens:
+
+- Draft acceptance increased from 55.42% to 56.57%.
+- Aggregate output throughput increased from 4.890 to 5.084 tokens per second.
+- Average request latency decreased from 88.64 to 84.45 seconds.
+
+All other model weights are unchanged.

model-00041-of-00041.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da5447af3c6c7d1cdbd8fa3a05ffa0e6c4c6f5e983fd04a2d794610b2cafd7aa
+oid sha256:b8f0616dbb1d82c16823f9d7f21e5e74bb2e178db271f4bedf2ad19b4e0bebd9
 size 5349720144