[OMNIML-4740] synth_support — pensieve-intern agent draft

Author: pensieve-intern

tools/launcher/common/service_utils.sh

@@ -18,8 +18,8 @@
 native_mpi_rank=$OMPI_COMM_WORLD_RANK
 native_mpi_local_rank=$OMPI_COMM_WORLD_LOCAL_RANK
 # Works with Slurm launching with `--mpi=pmix`
-mpi_rank=${PMIX_RANK:-$native_mpi_rank}
-mpi_local_rank=${PMIX_LOCAL_RANK:-$native_mpi_local_rank}
+mpi_rank=${PMIX_RANK:-${native_mpi_rank:-${SLURM_PROCID:-0}}}
+mpi_local_rank=${PMIX_LOCAL_RANK:-${native_mpi_local_rank:-${SLURM_LOCALID:-0}}}
 
 FAIL=0
 FAIL_EXIT=0
@@ -48,8 +48,23 @@ function report_result {
 }
 
 function util_install_extra_dep {
+    local _marker=/tmp/.nmm_extra_dep_installed
+    if [[ -f "$_marker" ]]; then
+        return 0
+    fi
     if [[ "$mpi_local_rank" -eq 0 ]]; then
         pip install diskcache
+        local _nvrx_dir
+        _nvrx_dir="$(mktemp -d)/nvidia-resiliency-ext"
+        git clone --depth 1 https://github.com/NVIDIA/nvidia-resiliency-ext "${_nvrx_dir}" \
+            && pip install "${_nvrx_dir}"
+        touch "$_marker"
+    else
+        local _waited=0
+        while [[ ! -f "$_marker" && $_waited -lt 600 ]]; do
+            sleep 1
+            _waited=$((_waited + 1))
+        done
     fi
 }
 

tools/launcher/examples/Kimi-K2.5/hf_offline_dflash.yaml

@@ -0,0 +1,38 @@
+# DFlash offline speculative decoding pipeline for Kimi-K2.5 (input model).
+# This YAML contains only task_0 (data synthesis) for synth_support stage.
+# Tasks 1-3 are added by downstream stages.
+
+job_name: Kimi-K2.5_DFlash_offline
+pipeline:
+  allow_to_fail: false
+  skip: false
+  note:
+
+  global_vars:
+    hf_model: /hf-local/moonshotai/Kimi-K2.6
+
+  # Step 1: Data synthesis via vLLM server
+  # Args before "--" go to vLLM server; args after "--" go to tools/query.py.
+  task_0:
+    script: common/vllm/query.sh
+    args:
+      - --model <<global_vars.hf_model>>
+      - --tensor-parallel-size 8
+      - --port 8000
+      - --host 0.0.0.0
+      - --trust_remote_code
+      - --enforce-eager
+      - --gpu-memory-utilization 0.95
+      - --max-model-len 4096
+      - --
+      - --data /hf-local/modelopt/Speculative-Decoding-Prompt-Samples
+      - --save /scratchspace/data
+    environment:
+      - HF_LOCAL: /hf-local
+      - VLLM_STARTUP_TIMEOUT: "1800"
+    slurm_config:
+      _factory_: "slurm_factory"
+      nodes: 1
+      ntasks_per_node: 1
+      gpus_per_node: 8
+      container: vllm/vllm-openai:latest