Upgrade containers

cluster_configs/example-local.yaml

@@ -15,9 +15,9 @@
 executor: local
 
 containers:
-  trtllm: nvcr.io/nvidia/tensorrt-llm/release:1.0.0
-  vllm: vllm/vllm-openai:v0.10.1.1
-  sglang: lmsysorg/sglang:v0.5.4
+  trtllm: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc1
+  vllm: dockerfile:dockerfiles/Dockerfile.vllm
+  sglang: lmsysorg/sglang:v0.5.8
   # dockerfile: for now can only specify relative to repo root
   megatron: dockerfile:dockerfiles/Dockerfile.megatron
   sandbox: dockerfile:dockerfiles/Dockerfile.sandbox

dockerfiles/Dockerfile.vllm

@@ -1,4 +1,4 @@
-FROM vllm/vllm-openai:v0.14.0
+FROM vllm/vllm-openai:v0.14.1
 RUN pip install "vllm[audio]"
 # Required by vLLM for Qwen-VL model family (runtime dependency, not directly imported)
 RUN pip install qwen-vl-utils

dockerfiles/README.md

@@ -28,12 +28,8 @@ set `DOCKER_PLATFORM=linux/arm64` for the build script described above.
 
 ## Building trtllm image
 
-We directly use official `nvcr.io/nvidia/tensorrt-llm/release:1.0.0` image for both amd64 and arm64.
+We directly use official `nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc1` image.
 
 ## Building sglang image
 
-We directly use official `lmsysorg/sglang:v0.5.4` image.
-
-## Building vllm image
-
-We use official `vllm/vllm-openai:v0.10.2` image with the additional `vllm[audio]` dependencies.
+We directly use official `lmsysorg/sglang:v0.5.8` image.

nemo_skills/__init__.py

@@ -16,9 +16,9 @@
 
 # only used in ns setup command to initialize with defaults
 _containers = {
-    "trtllm": "nvcr.io/nvidia/tensorrt-llm/release:1.0.0",
+    "trtllm": "nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc1",
     "vllm": "dockerfile:dockerfiles/Dockerfile.vllm",
-    "sglang": "lmsysorg/sglang:v0.5.4",
+    "sglang": "lmsysorg/sglang:v0.5.8",
     "megatron": "dockerfile:dockerfiles/Dockerfile.megatron",
     "sandbox": "dockerfile:dockerfiles/Dockerfile.sandbox",
     "nemo-skills": "dockerfile:dockerfiles/Dockerfile.nemo-skills",

tests/gpu-tests/test-local.yaml

@@ -15,9 +15,9 @@
 executor: local
 
 containers:
-  trtllm: nvcr.io/nvidia/tensorrt-llm/release:1.0.0
+  trtllm: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc1
   vllm: dockerfile:dockerfiles/Dockerfile.vllm
-  sglang: lmsysorg/sglang:v0.5.4
+  sglang: lmsysorg/sglang:v0.5.8
   sandbox: dockerfile:dockerfiles/Dockerfile.sandbox
   nemo-skills: dockerfile:dockerfiles/Dockerfile.nemo-skills
   megatron: dockerfile:dockerfiles/Dockerfile.megatron

tests/slurm-tests/qwen3_4b_evals/check_results.py

@@ -24,13 +24,13 @@
 TOOLCALLING_METRIC_RANGES = {
     ("overall_accuracy", "accuracy"): (61.0, 67.0),
     ("overall_non_live", "accuracy"): (84.0, 90.0),
-    ("non_live_ast", "accuracy"): (85.0, 92.0),
+    ("non_live_ast", "accuracy"): (84.0, 92.0),
     ("non_live_irrelevance", "accuracy"): (79.0, 86.0),
     ("overall_live", "accuracy"): (76.0, 83.0),
     ("live_ast", "accuracy"): (79.0, 86.0),
     ("live_irrelevance", "accuracy"): (73.0, 80.0),
-    ("live_relevance", "accuracy"): (70.0, 90.0),  # unusually high variance
-    ("overall_multi_turn", "accuracy"): (20.0, 30.0),
+    ("live_relevance", "accuracy"): (70.0, 100.0),  # unusually high variance
+    ("overall_multi_turn", "accuracy"): (20.0, 33.0),
 }