truss 0.11.13rc3__tar.gz → 0.11.13rc4__tar.gz

{truss-0.11.13rc3 → truss-0.11.13rc4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: truss
-Version: 0.11.13rc3
+Version: 0.11.13rc4
 Summary: A seamless bridge from model development to model delivery
 Project-URL: Repository, https://github.com/basetenlabs/truss
 Project-URL: Homepage, https://truss.baseten.co
@@ -37,7 +37,7 @@ Requires-Dist: rich<14,>=13.4.2
 Requires-Dist: ruff>=0.4.8
 Requires-Dist: tenacity>=8.0.1
 Requires-Dist: tomlkit>=0.13.2
-Requires-Dist: truss-transfer<0.0.40,>=0.0.36
+Requires-Dist: truss-transfer<0.0.40,>=0.0.37
 Requires-Dist: watchfiles<0.20,>=0.19.0
 Description-Content-Type: text/markdown
 

{truss-0.11.13rc3 → truss-0.11.13rc4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "truss"
-version = "0.11.13rc3"
+version = "0.11.13rc4"
 description = "A seamless bridge from model development to model delivery"
 authors = [
     { name = "Pankaj Gupta", email = "no-reply@baseten.co" },
@@ -43,7 +43,7 @@ dependencies = [
     "ruff>=0.4.8",
     "tenacity>=8.0.1",
     "watchfiles>=0.19.0,<0.20",
-    "truss-transfer>=0.0.36,<0.0.40",
+    "truss-transfer>=0.0.37,<0.0.40",
 ]
 
 [project.urls]

{truss-0.11.13rc3 → truss-0.11.13rc4}/truss/base/trt_llm_config.py

@@ -68,6 +68,7 @@ class TrussTRTLLMQuantizationType(str, Enum):
     FP8_KV = "fp8_kv"
     FP4 = "fp4"
     FP4_KV = "fp4_kv"
+    FP4_MLP_ONLY = "fp4_mlp_only"
 
 
 class TrussTRTLLMPluginConfiguration(PydanticTrTBaseModel):
@@ -713,7 +714,9 @@ def trt_llm_common_validation(config: "TrussConfig"):
             "accelerators or newer (CUDA_COMPUTE>=89)"
         )
     elif trt_llm_config.build.quantization_type in [
-        TrussTRTLLMQuantizationType.FP4
+        TrussTRTLLMQuantizationType.FP4,
+        TrussTRTLLMQuantizationType.FP4_KV,
+        TrussTRTLLMQuantizationType.FP4_MLP_ONLY,
     ] and config.resources.accelerator.accelerator in [
         truss_config.Accelerator.H100,
         truss_config.Accelerator.L4,

{truss-0.11.13rc3 → truss-0.11.13rc4}/truss/templates/server.Dockerfile.jinja

@@ -69,7 +69,7 @@ COPY --chown={{ default_owner }} ./{{ config.data_dir }} ${APP_HOME}/data
 
 {%- if model_cache_v2 %}
 {# v0.0.9, keep synced with server_requirements.txt #}
-RUN curl -sSL --fail --retry 5 --retry-delay 2 -o /usr/local/bin/truss-transfer-cli https://github.com/basetenlabs/truss/releases/download/v0.11.12rc4/truss-transfer-cli-v0.11.12rc4-linux-x86_64-unknown-linux-musl
+RUN curl -sSL --fail --retry 5 --retry-delay 2 -o /usr/local/bin/truss-transfer-cli https://github.com/basetenlabs/truss/releases/download/v0.11.13rc3/truss-transfer-cli-v0.11.13rc3-linux-x86_64-unknown-linux-musl
 RUN chmod +x /usr/local/bin/truss-transfer-cli
 RUN mkdir /static-bptr
 RUN echo "hash {{model_cache_hash}}"

{truss-0.11.13rc3 → truss-0.11.13rc4}/truss/templates/shared/lazy_data_resolver.py

@@ -75,14 +75,21 @@ class TrussTransferStats:
         if not PROMETHEUS_AVAILABLE:
             return
         global METRICS_REGISTERED
-        if not METRICS_REGISTERED:
+
+        if METRICS_REGISTERED:
+            logging.info(
+                "Model cache metrics already registered, skipping."
+            )  # this should never happen
+            return
+        else:
             # Ensure metrics are only registered once
             METRICS_REGISTERED = True
 
-        # Define metrics with model_cache label
+        # Define metrics with model_cache prefix
         manifest_size_gauge = Gauge(
             "model_cache_manifest_size_bytes", "Total manifest size in bytes"
         )
+        # histograms have intentially wide buckets to capture a variety of download times
         download_time_histogram = Histogram(
             "model_cache_download_time_seconds",
             "Total download time in seconds",