Add Mistral Small 4 H100 baseline and optimized configs

models/mistral-small-4-119b-2603/metadata.yaml

@@ -0,0 +1,23 @@
+apiVersion: catalog.swiss/v1
+name: mistral-small-4-119b-2603
+displayName: Mistral-Small-4-119B-2603
+description: Mistral-Small-4-119B-2603 on 2 H100 GPUs. Configs require the
+  measured tokenizer at /tokenizer.
+family: mistral
+tags:
+  - chat
+  - reasoning
+  - tool-use
+source:
+  hf: mistralai/Mistral-Small-4-119B-2603
+tuning:
+  - version: 1.0.0
+    baseline: sglang-tp2-h100-baseline
+    optimized: sglang-tp2-h100-optimized
+    uplift: 52.5
+    workloads:
+      - name: 50k + 1.5k
+        uplift: 52.5
+      - name: 8k + 1k
+        uplift: 32.1
+    report: mistral-small-4-119b-2603-h100-report.html

models/mistral-small-4-119b-2603/mistral-small-4-119b-2603-1.0.0.yaml

@@ -0,0 +1,151 @@
+# Serving configs measured, and where marked tuned, by LLM AutoTune.
+apiVersion: catalog.swiss/v1
+name: mistral-small-4-119b-2603
+version: 1.0.0
+servedName: Mistral-Small-4-119B-2603
+variants:
+  - id: sglang-tp2-h100-baseline
+    description: 'Baseline: TP2 on two H100 GPUs, single node.'
+    engine: sglang
+    chart:
+      name: sglang
+      version: ">=0.7.1"
+    image:
+      repository: harbor.4pd.io/hardcore-tech/sglang
+      tag: v0.5.19-cu129
+    requires:
+      gpus: 2
+      topology: single-node
+      gpuProduct:
+        - NVIDIA-H100-80GB-HBM3
+    values:
+      model:
+        hostPathType: Directory
+        contextLength: ''
+      modelCheck:
+        enabled: true
+        requiredGlobs:
+          - config.json
+          - '*.safetensors'
+      extraArgs:
+        - --load-format=mistral
+        - --model-config-parser=mistral
+        - --reasoning-parser=mistral
+        - --tokenizer-path=/tokenizer # Mount the measured tokenizer at /tokenizer via the site profile.
+        - --tp-size=2
+      env:
+        - name: NVIDIA_DISABLE_REQUIRE
+          value: '1'
+        - name: PYTORCH_ALLOC_CONF
+          value: expandable_segments:True
+      volumes:
+        - name: dot-cache
+          emptyDir:
+            sizeLimit: 10Gi
+        - name: dshm
+          emptyDir:
+            medium: Memory
+            sizeLimit: 32Gi
+      volumeMounts:
+        - name: dot-cache
+          mountPath: /root/.cache
+        - name: dshm
+          mountPath: /dev/shm
+      resources:
+        requests:
+          ephemeral-storage: 16Gi
+        limits:
+          ephemeral-storage: 16Gi
+      progressDeadlineSeconds: 7200
+      startupProbe:
+        httpGet:
+          path: /health
+          port: http
+        periodSeconds: 30
+        timeoutSeconds: 10
+        failureThreshold: 180
+      terminationGracePeriodSeconds: 3600
+      lifecycle:
+        preStop:
+          drainSeconds: 600
+          pollIntervalSeconds: 5
+        forceShutdown: true
+        preStopKill: true
+      hangWatcher:
+        enabled: true
+  - id: sglang-tp2-h100-optimized
+    default: true
+    description: 'Tuned by LLM AutoTune: TP2 on two H100 GPUs, single node.
+      Normalized throughput within the SLO vs the baseline variant: +52.5% (50k
+      + 1.5k), +32.1% (8k + 1k).'
+    engine: sglang
+    chart:
+      name: sglang
+      version: ">=0.7.1"
+    image:
+      repository: harbor.4pd.io/hardcore-tech/sglang
+      tag: v0.5.19-cu129
+      digest: sha256:b7a3051b9e15939a5afc74de3564a72fd5b0b42bef80bbaacea9649d09f63d76
+    requires:
+      gpus: 2
+      topology: single-node
+      gpuProduct:
+        - NVIDIA-H100-80GB-HBM3
+    values:
+      model:
+        hostPathType: Directory
+        contextLength: ''
+      modelCheck:
+        enabled: true
+        requiredGlobs:
+          - config.json
+          - '*.safetensors'
+      extraArgs:
+        - --chunked-prefill-size=8192
+        - --load-format=mistral
+        - --mem-fraction-static=0.86
+        - --model-config-parser=mistral
+        - --reasoning-parser=mistral
+        - --schedule-conservativeness=1.3
+        - --tokenizer-path=/tokenizer # Mount the measured tokenizer at /tokenizer via the site profile.
+        - --tp-size=2
+      env:
+        - name: NVIDIA_DISABLE_REQUIRE
+          value: '1'
+        - name: PYTORCH_ALLOC_CONF
+          value: expandable_segments:True
+      volumes:
+        - name: dot-cache
+          emptyDir:
+            sizeLimit: 10Gi
+        - name: dshm
+          emptyDir:
+            medium: Memory
+            sizeLimit: 32Gi
+      volumeMounts:
+        - name: dot-cache
+          mountPath: /root/.cache
+        - name: dshm
+          mountPath: /dev/shm
+      resources:
+        requests:
+          ephemeral-storage: 16Gi
+        limits:
+          ephemeral-storage: 16Gi
+      progressDeadlineSeconds: 7200
+      startupProbe:
+        httpGet:
+          path: /health
+          port: http
+        periodSeconds: 30
+        timeoutSeconds: 10
+        failureThreshold: 180
+      terminationGracePeriodSeconds: 3600
+      lifecycle:
+        preStop:
+          drainSeconds: 600
+          pollIntervalSeconds: 5
+        forceShutdown: true
+        preStopKill: true
+      hangWatcher:
+        enabled: true