Enable Strix Halo profiling support

.gitignore

@@ -1,6 +1,17 @@
-**/*cache*/
+# venv
+venv/
+.venv/
+
+# python cache
 __pycache__/
+*.pyc
+
+# generic caches
+**/*cache*/
 
-*.*/
+#files
+images/
 
-Dockerfile
+# logs (optional but recommended)
+.logs/
+.config/

README.md

@@ -30,6 +30,16 @@ Purpose:
 * Preserve downloaded models between runs
 * Avoid repeated downloads across container executions
 
+### `.cache/vllm/` and `.cache/triton/`
+
+These directories are mounted into Docker containers to persist vLLM and Triton compile caches.
+
+Purpose:
+
+* Reduce repeated `torch.compile` and CUDA graph capture time across runs
+* Make Strix Halo profiling runs closer to steady-state behavior after the first cold start
+* Improve repeatability when comparing baselines across models
+
 ---
 
 ### `.config/`
@@ -145,7 +155,13 @@ Run directly on the host system:
 * `generate_gpu_yaml.sh`
   Helper script to auto-generate a `gpus.yaml` template
 
-  **NOTE**: *This script is currently defunct and you will have to create your own .yaml config based on the one provided in the README and your own setup*
+  Usage:
+
+  ```bash
+  scripts/host/generate_gpu_yaml.sh
+  ```
+
+  This generates `.config/gpus.yaml` from the local `/dev/dri/renderD*` devices.
 
 * Future host utilities will also live here
 
@@ -209,3 +225,21 @@ Notes:
 * Improved automation and validation
 * Expanded profiling support
 * Potential non-Linux support
+
+---
+
+## Portability Notes
+
+* Host scripts now resolve workspace paths from the repository root, so running commands from different working directories is supported.
+* Use `disabled_on` in `yaml/models.yaml` to prevent known-problematic models on specific devices.
+
+Example:
+
+```yaml
+models:
+  - name: Qwen/Qwen3-VL-4B-Instruct
+    type: multimodal
+    script: qwen_vl.py
+    disabled_on:
+      - Strix Halo
+```

scripts/host/docker_tool.py

@@ -40,6 +40,10 @@ def prepare_env():
 
 def run_container(args, script_args):
     env = prepare_env()
+    hf_cache_dir = Path(args.hf_cache_dir)
+    if not hf_cache_dir.is_absolute():
+        hf_cache_dir = (ROOT_DIR / hf_cache_dir).resolve()
+
     cmd = [
         "docker",
         "run",
@@ -81,34 +85,50 @@ def run_container(args, script_args):
 
     # huggingface cache dir
     hf_cache_container = "/root/.cache/huggingface"
-    Path(args.hf_cache_dir).mkdir(parents=True, exist_ok=True)
-    cmd.extend(["-v", f"{args.hf_cache_dir}:{hf_cache_container}"])
-    print(f"Mounting HuggingFace cache: {args.hf_cache_dir} -> {hf_cache_container}")
+    hf_cache_dir.mkdir(parents=True, exist_ok=True)
+    cmd.extend(["-v", f"{hf_cache_dir}:{hf_cache_container}"])
+    print(f"Mounting HuggingFace cache: {hf_cache_dir} -> {hf_cache_container}")
+
+    # vLLM and Triton compile caches
+    cache_mounts = {
+        "/root/.cache/vllm": ROOT_DIR / ".cache" / "vllm",
+        "/root/.cache/triton": ROOT_DIR / ".cache" / "triton",
+    }
+    for container_cache_dir, host_cache_dir in cache_mounts.items():
+        host_cache_dir.mkdir(parents=True, exist_ok=True)
+        cmd.extend(["-v", f"{host_cache_dir}:{container_cache_dir}"])
+        print(f"Mounting cache: {host_cache_dir} -> {container_cache_dir}")
 
     # local container scripts dir
     scripts_container = str(container_workspace / "scripts")
-    cmd.extend(["-v", f"./scripts/container:{scripts_container}"])
-    print(f"Mounting ./scripts/container -> {scripts_container}")
+    host_scripts_dir = ROOT_DIR / "scripts" / "container"
+    cmd.extend(["-v", f"{host_scripts_dir}:{scripts_container}"])
+    print(f"Mounting {host_scripts_dir} -> {scripts_container}")
 
     # local prompts dir
     prompts_container = str(container_workspace / "prompts")
-    cmd.extend(["-v", f"./prompts:{prompts_container}"])
-    print(f"Mounting ./prompts -> {prompts_container}")
+    host_prompts_dir = ROOT_DIR / "prompts"
+    cmd.extend(["-v", f"{host_prompts_dir}:{prompts_container}"])
+    print(f"Mounting {host_prompts_dir} -> {prompts_container}")
 
     # logs dir
     logs_container = str(container_workspace / "logs")
-    cmd.extend(["-v", f"./.logs:{logs_container}"])
-    print(f"Mounting ./.logs -> {logs_container}")
+    host_logs_dir = ROOT_DIR / ".logs"
+    host_logs_dir.mkdir(parents=True, exist_ok=True)
+    cmd.extend(["-v", f"{host_logs_dir}:{logs_container}"])
+    print(f"Mounting {host_logs_dir} -> {logs_container}")
 
     # images dir
     images_container = str(container_workspace / "images")
-    cmd.extend(["-v", f"./images:{images_container}"])
-    print(f"Mounting ./images -> {images_container}")
+    host_images_dir = ROOT_DIR / "images"
+    cmd.extend(["-v", f"{host_images_dir}:{images_container}"])
+    print(f"Mounting {host_images_dir} -> {images_container}")
 
     # yaml dir
     yaml_container = str(container_workspace / "yaml")
-    cmd.extend(["-v", f"./yaml:{yaml_container}"])
-    print(f"Mounting ./yaml -> {yaml_container}")
+    host_yaml_dir = ROOT_DIR / "yaml"
+    cmd.extend(["-v", f"{host_yaml_dir}:{yaml_container}"])
+    print(f"Mounting {host_yaml_dir} -> {yaml_container}")
 
     shell_cmd = []
 

scripts/host/generate_gpu_yaml.sh

@@ -1,33 +1,36 @@
 #!/usr/bin/env bash
 
-echo "THE SCRIPT generate_gpu_yaml.sh IS CURRENTLY DEFUNCT"
-echo "GENERATE THE FILE MANUALLY"
-echo "EXAMPLE:"
-echo "gpus:
-        0:
-          name: Radeon RX 7900 XTX
-          device: /dev/dri/renderD128
-        1:
-          name: AMD Radeon RX 9070 XT
-          device: /dev/dri/renderD129
-        2:
-          name: AMD Radeon RX 6700 XT
-          device: /dev/dri/renderD130"
-# TODO: ENABLE USING ROCMINFO
-
-exit 1
-
-echo "gpus:"
-
-i=0
-for dev in /dev/dri/renderD*; do
-  name=$(udevadm info --query=property --name="$dev" \
-    | grep ID_MODEL= \
-    | cut -d= -f2)
-
-  echo "  gpu$i:"
-  echo "    name: ${name:-unknown}"
-  echo "    device: $dev"
-
-  ((i++))
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "$0")/../.." && pwd)"
+OUT_FILE="${ROOT_DIR}/.config/gpus.yaml"
+
+mkdir -p "${ROOT_DIR}/.config"
+
+mapfile -t render_devices < <(ls /dev/dri/renderD* 2>/dev/null | sort || true)
+
+if [[ ${#render_devices[@]} -eq 0 ]]; then
+  echo "No /dev/dri/renderD* devices found; cannot generate ${OUT_FILE}" >&2
+  exit 1
+fi
+
+echo "gpus:" > "${OUT_FILE}"
+
+for dev in "${render_devices[@]}"; do
+  name=$(udevadm info --query=property --name="$dev" 2>/dev/null \
+    | awk -F= '/^ID_MODEL_FROM_DATABASE=/{print $2; found=1; exit} /^ID_MODEL=/{if (!found) print $2}' \
+    | sed 's/_/ /g' \
+    || true)
+
+  if [[ -z "${name}" ]]; then
+    name="${dev##*/}"
+  fi
+
+  cat >> "${OUT_FILE}" <<EOF
+  - name: "${name}"
+    device: "${dev}"
+    disabled: false
+EOF
 done
+
+echo "Generated ${OUT_FILE} with ${#render_devices[@]} GPU device(s)."

scripts/host/orchestrator.py

@@ -14,6 +14,7 @@
 from pathlib import Path
 import yaml
 import os
+import sys
 
 PROJECT_ROOT = Path(__file__).parent.parent.parent
 
@@ -32,7 +33,11 @@ def prepare_tokens():
 def parse_gpus():
     file_path = PROJECT_ROOT / ".config" / "gpus.yaml"
     if not file_path.exists():
-        subprocess.run(["generate_gpu_yaml.sh"], check=True)
+        subprocess.run(
+            [str(PROJECT_ROOT / "scripts" / "host" / "generate_gpu_yaml.sh")],
+            check=True,
+            cwd=PROJECT_ROOT,
+        )
     with file_path.open("r") as f:
         return [
             gpu for gpu in yaml.safe_load(f)["gpus"] if not gpu.get("disabled", False)
@@ -102,7 +107,8 @@ def run(docker_image, num_procs, script, duration, iterations, models_filter):
                 ]
                 result = subprocess.run(
                     [
-                        "scripts/host/docker_tool.py",
+                        sys.executable,
+                        str(PROJECT_ROOT / "scripts" / "host" / "docker_tool.py"),
                         "run",
                         "--image-name",
                         docker_image,
@@ -116,6 +122,7 @@ def run(docker_image, num_procs, script, duration, iterations, models_filter):
                     ]
                     + script_args,
                     env=os.environ.copy() | env,
+                    cwd=PROJECT_ROOT,
                 )
 
 

yaml/env_vars.yaml

@@ -26,3 +26,4 @@ env_vars:
   - NCCL_NVLS_ENABLE
   - TORCH_NCCL_AVOID_RECORD_STREAMS
   - PYTORCH_CUDA_ALLOC_CONF
+  - PYTORCH_ALLOC_CONF

yaml/models.yaml

@@ -15,9 +15,11 @@ models:
   - name: Qwen/Qwen3-VL-4B-Instruct
     type: multimodal
     script: qwen_vl.py
+    disabled_on:
+      - Strix Halo
     env:
       VLLM_WORKER_MULTIPROC_METHOD: 'spawn'
-      MAX_MODEL_LEN: '20000'
+      MAX_MODEL_LEN: '2048'
       SP_TEMPERATURE: '0.0'
       SP_MAX_TOKENS: '1024'
 #  - name: deepseek-ai/DeepSeek-OCR