add deepseek v4 attempt

Author: robmsmt

examples/clariden/cli/deepseek-ai/DeepSeek-V4-Pro-Base-sglang.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+# 1.6T params, fp8 block-quantised. 8 nodes x TP32 (32 GH200s).
+# Requires sglang with DeepseekV4ForCausalLM support and transformers >= 4.57.1.
+sml advanced \
+  --firecrest-system clariden \
+  --partition normal \
+  --slurm-nodes-per-replica 8 \
+  --slurm-time 3:00:00 \
+  --serving-framework sglang \
+  --slurm-environment src/swiss_ai_model_launch/assets/envs/sglang.toml \
+  --framework-args "--model-path /capstor/store/cscs/swissai/infra01/hf_models/models/deepseek-ai/DeepSeek-V4-Pro-Base \
+    --served-model-name deepseek-ai/DeepSeek-V4-Pro-Base-$(whoami) \
+    --tp-size 32 \
+    --host 0.0.0.0 \
+    --trust-remote-code \
+    --mem-fraction-static 0.85 \
+    --enable-metrics"

examples/clariden/cli/deepseek-ai/DeepSeek-V4-Pro-sglang.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+# Post-trained DeepSeek-V4-Pro, MXFP4-quantised (~806 GB on disk).
+# 4 nodes x TP16 mirrors V3.1 density. Requires sglang with MXFP4 dequant
+# kernels and DeepseekV4ForCausalLM support; transformers >= 4.57.1.
+sml advanced \
+  --firecrest-system clariden \
+  --partition normal \
+  --slurm-nodes-per-replica 4 \
+  --serving-framework sglang \
+  --slurm-time 3:00:00 \
+  --slurm-environment src/swiss_ai_model_launch/assets/envs/sglang.toml \
+  --framework-args "--model-path /capstor/store/cscs/swissai/infra01/hf_models/models/deepseek-ai/DeepSeek-V4-Pro \
+    --served-model-name deepseek-ai/DeepSeek-V4-Pro-$(whoami) \
+    --tp-size 16 \
+    --host 0.0.0.0 \
+    --trust-remote-code \
+    --mem-fraction-static 0.85 \
+    --enable-metrics"