feat(strixhalo): add amdsmi sysfs shim to bypass glibc 2.38 requirement

The native amdsmi from ROCm 7.1 requires libamd_smi.so linked against
glibc 2.38 (Ubuntu 24.04), but the Ray base image is Ubuntu 22.04
(glibc 2.35). This caused vLLM to fail ROCm platform detection with
'No module named amdsmi' / GLIBC_2.38 not found errors.

Solution: Pure-Python amdsmi shim that reads GPU info from sysfs
(/sys/class/drm/*) instead of the native library. Provides the full
API surface used by both vLLM (platform detection, device info) and
PyTorch (device counting, memory/power/temp monitoring).

Tested in-container: vLLM detects RocmPlatform, PyTorch sees GPU
(Radeon 8060S, 128GB, HIP 7.3), DeviceConfig resolves to 'cuda'.

Changes:
- Add amdsmi-shim/ package with sysfs-backed implementation
- Update Dockerfile to install shim after vLLM/torch
- Add amdsmi-shim/ to .dockerignore explicit includes

.dockerignore

@@ -42,3 +42,4 @@ Makefile
 # Don't ignore these (explicitly include)
 !ray-serve/
 !dockerfiles/
+!amdsmi-shim/

amdsmi-shim/amdsmi/__init__.py

@@ -0,0 +1,335 @@
+"""
+Pure-Python amdsmi shim using sysfs for ROCm platform detection.
+
+Replaces the native amdsmi package (which requires libamd_smi.so linked
+against glibc 2.38) with a lightweight implementation that reads GPU
+information from sysfs (/sys/class/drm/*).
+
+This is needed because the Ray base image (Ubuntu 22.04, glibc 2.35) cannot
+load the ROCm 7.1 vendor libamd_smi.so (built for Ubuntu 24.04, glibc 2.38).
+
+Provides the subset of the amdsmi API used by:
+  - vLLM (platform detection, device info, topology)
+  - PyTorch (device counting, memory/power/temp monitoring)
+"""
+
+from __future__ import annotations
+
+import glob
+import logging
+import os
+from enum import IntEnum
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Exception
+# ---------------------------------------------------------------------------
+
+
+class AmdSmiException(Exception):
+    """Shim for amdsmi.AmdSmiException."""
+
+    def __init__(self, message: str = "", err_code: int = -1):
+        super().__init__(message)
+        self.err_code = err_code
+
+
+# ---------------------------------------------------------------------------
+# Enums used by PyTorch monitoring helpers
+# ---------------------------------------------------------------------------
+
+
+class AmdSmiTemperatureType(IntEnum):
+    EDGE = 0
+    JUNCTION = 1
+    VRAM = 2
+
+
+class AmdSmiTemperatureMetric(IntEnum):
+    CURRENT = 0
+    MAX = 1
+    MIN = 2
+
+
+class AmdSmiClkType(IntEnum):
+    GFX = 0
+    MEM = 1
+
+
+# ---------------------------------------------------------------------------
+# sysfs helpers
+# ---------------------------------------------------------------------------
+
+_DRM_BASE = "/sys/class/drm"
+
+
+def _read_sysfs(path: str, default: str = "") -> str:
+    """Read a sysfs file, returning *default* on any error."""
+    try:
+        return Path(path).read_text().strip()
+    except (OSError, IOError):
+        return default
+
+
+def _read_sysfs_int(path: str, default: int = 0) -> int:
+    val = _read_sysfs(path)
+    try:
+        if val.startswith("0x"):
+            return int(val, 16)
+        return int(val)
+    except (ValueError, TypeError):
+        return default
+
+
+def _discover_cards() -> list[str]:
+    """Return sorted list of card directory names, e.g. ['card0', 'card1']."""
+    cards = []
+    for entry in sorted(glob.glob(os.path.join(_DRM_BASE, "card[0-9]*"))):
+        name = os.path.basename(entry)
+        # Exclude render nodes and connector entries like card0-DP-1
+        if name.startswith("card") and "-" not in name:
+            vendor = _read_sysfs(os.path.join(entry, "device", "vendor"))
+            # 0x1002 = AMD
+            if vendor == "0x1002":
+                cards.append(name)
+    return cards
+
+
+def _find_hwmon(card: str) -> str | None:
+    """Return the first hwmon directory path for a card, or None."""
+    hwmon_base = os.path.join(_DRM_BASE, card, "device", "hwmon")
+    entries = sorted(glob.glob(os.path.join(hwmon_base, "hwmon*")))
+    return entries[0] if entries else None
+
+
+# ---------------------------------------------------------------------------
+# Handle management
+# ---------------------------------------------------------------------------
+
+# Each "handle" is just a dict with cached paths
+_handles: list[dict[str, Any]] = []
+_initialized = False
+
+
+def amdsmi_init() -> None:
+    """Initialise the shim by discovering AMD GPUs via sysfs."""
+    global _handles, _initialized
+    if _initialized:
+        return
+    cards = _discover_cards()
+    _handles = []
+    for card in cards:
+        dev = os.path.join(_DRM_BASE, card, "device")
+        _handles.append(
+            {
+                "card": card,
+                "dev": dev,
+                "hwmon": _find_hwmon(card),
+                "vendor": _read_sysfs(os.path.join(dev, "vendor")),
+                "device_id": _read_sysfs(os.path.join(dev, "device")),
+            }
+        )
+    _initialized = True
+    logger.debug("amdsmi-shim: discovered %d AMD GPU(s) via sysfs", len(_handles))
+
+
+def amdsmi_shut_down() -> None:
+    """Reset internal state."""
+    global _handles, _initialized
+    _handles = []
+    _initialized = False
+
+
+def amdsmi_get_processor_handles() -> list[dict[str, Any]]:
+    """Return a list of opaque handles (one per GPU)."""
+    if not _initialized:
+        raise AmdSmiException("amdsmi not initialised", err_code=1)
+    return list(_handles)
+
+
+# ---------------------------------------------------------------------------
+# Device info  (used by vLLM get_device_name + PyTorch UUID helpers)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_gpu_asic_info(handle: dict[str, Any]) -> dict[str, str]:
+    """
+    Return ASIC information for a GPU handle.
+
+    Keys returned: device_id, market_name, asic_serial, vendor_id, rev_id
+    """
+    dev = handle["dev"]
+    device_id_hex = _read_sysfs(os.path.join(dev, "device"), "0x0000")
+    # Read marketing / product name (may not exist on all kernels)
+    market = (
+        _read_sysfs(os.path.join(dev, "product_name"))
+        or _read_sysfs(os.path.join(dev, "marketing_name"))
+        or f"AMD GPU {device_id_hex}"
+    )
+    # Unique ID for serial (not always available)
+    serial = _read_sysfs(os.path.join(dev, "unique_id"), "0000000000000000")
+    return {
+        "device_id": device_id_hex,
+        "market_name": market,
+        "asic_serial": serial,
+        "vendor_id": _read_sysfs(os.path.join(dev, "vendor"), "0x1002"),
+        "rev_id": _read_sysfs(os.path.join(dev, "revision"), "0x00"),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Topology  (used by vLLM is_fully_connected)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_topo_get_link_type(
+    handle: dict[str, Any], peer_handle: dict[str, Any]
+) -> dict[str, int]:
+    """
+    Return link topology between two GPUs.
+
+    For single-GPU systems this is essentially a no-op.
+    type: 2 = XGMI, 1 = PCIe;  hops = number of hops.
+    """
+    if handle is peer_handle or handle["card"] == peer_handle["card"]:
+        return {"type": 2, "hops": 0}  # same device
+    # Default: not XGMI-connected
+    return {"type": 1, "hops": 2}
+
+
+# ---------------------------------------------------------------------------
+# Memory  (used by PyTorch _get_amdsmi_device_memory_used)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_gpu_vram_usage(handle: dict[str, Any]) -> dict[str, int]:
+    """
+    Return VRAM usage in megabytes.
+
+    Keys: vram_used, vram_total
+    """
+    dev = handle["dev"]
+    total_bytes = _read_sysfs_int(os.path.join(dev, "mem_info_vram_total"))
+    used_bytes = _read_sysfs_int(os.path.join(dev, "mem_info_vram_used"))
+    return {
+        "vram_used": used_bytes // (1024 * 1024),
+        "vram_total": total_bytes // (1024 * 1024),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Activity  (used by PyTorch _get_amdsmi_utilization / memory_usage)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_gpu_activity(handle: dict[str, Any]) -> dict[str, int]:
+    """
+    Return GPU and memory controller activity percentages.
+
+    Keys: gfx_activity, umc_activity, mm_activity
+    """
+    dev = handle["dev"]
+    gfx = _read_sysfs_int(os.path.join(dev, "gpu_busy_percent"))
+    # mem_busy_percent is not always available
+    umc = _read_sysfs_int(os.path.join(dev, "mem_busy_percent"))
+    return {"gfx_activity": gfx, "umc_activity": umc, "mm_activity": 0}
+
+
+# ---------------------------------------------------------------------------
+# Power  (used by PyTorch _get_amdsmi_power_draw)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_power_info(handle: dict[str, Any]) -> dict[str, int]:
+    """
+    Return power info in watts.
+
+    Keys: average_socket_power, current_socket_power
+    """
+    hwmon = handle.get("hwmon")
+    if not hwmon:
+        return {"average_socket_power": 0, "current_socket_power": 0}
+    # sysfs reports in microwatts
+    avg_uw = _read_sysfs_int(os.path.join(hwmon, "power1_average"))
+    cur_uw = _read_sysfs_int(os.path.join(hwmon, "power1_input"), avg_uw)
+    return {
+        "average_socket_power": avg_uw // 1_000_000,
+        "current_socket_power": cur_uw // 1_000_000,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Temperature  (used by PyTorch _get_amdsmi_temperature)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_temp_metric(
+    handle: dict[str, Any],
+    sensor_type: AmdSmiTemperatureType = AmdSmiTemperatureType.JUNCTION,
+    metric: AmdSmiTemperatureMetric = AmdSmiTemperatureMetric.CURRENT,
+) -> int:
+    """
+    Return temperature in millidegrees Celsius.
+
+    PyTorch uses JUNCTION / CURRENT and the result is divided by 1000.
+    """
+    hwmon = handle.get("hwmon")
+    if not hwmon:
+        return 0
+    # Map sensor type to sysfs file: temp1=edge, temp2=junction, temp3=vram
+    idx = {
+        AmdSmiTemperatureType.EDGE: 1,
+        AmdSmiTemperatureType.JUNCTION: 2,
+        AmdSmiTemperatureType.VRAM: 3,
+    }.get(sensor_type, 1)
+    suffix = {
+        AmdSmiTemperatureMetric.CURRENT: "input",
+        AmdSmiTemperatureMetric.MAX: "max",
+        AmdSmiTemperatureMetric.MIN: "min",
+    }.get(metric, "input")
+    # sysfs already reports millidegrees
+    return _read_sysfs_int(os.path.join(hwmon, f"temp{idx}_{suffix}"))
+
+
+# ---------------------------------------------------------------------------
+# Clock  (used by PyTorch _get_amdsmi_clock_rate via vLLM reference)
+# ---------------------------------------------------------------------------
+
+
+def amdsmi_get_clock_info(
+    handle: dict[str, Any],
+    clk_type: AmdSmiClkType = AmdSmiClkType.GFX,
+) -> dict[str, int]:
+    """
+    Return clock info in MHz.
+
+    Keys: clk (current), max_clk, min_clk
+    """
+    dev = handle["dev"]
+    sysfs_file = {
+        AmdSmiClkType.GFX: "pp_dpm_sclk",
+        AmdSmiClkType.MEM: "pp_dpm_mclk",
+    }.get(clk_type, "pp_dpm_sclk")
+
+    raw = _read_sysfs(os.path.join(dev, sysfs_file))
+    current_mhz = 0
+    min_mhz = 0
+    max_mhz = 0
+    for line in raw.splitlines():
+        # Lines look like: "0: 600Mhz *"
+        parts = line.split()
+        if len(parts) >= 2:
+            try:
+                freq = int(parts[1].replace("Mhz", "").replace("MHz", ""))
+            except ValueError:
+                continue
+            if min_mhz == 0 or freq < min_mhz:
+                min_mhz = freq
+            if freq > max_mhz:
+                max_mhz = freq
+            if "*" in line:
+                current_mhz = freq
+    return {"clk": current_mhz, "max_clk": max_mhz, "min_clk": min_mhz}

amdsmi-shim/pyproject.toml

@@ -0,0 +1,13 @@
+[build-system]
+requires = ["setuptools>=64"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "amdsmi"
+version = "0.1.0"
+description = "Lightweight sysfs-based amdsmi shim for ROCm platform detection without libamd_smi.so"
+requires-python = ">=3.8"
+license = "MIT"
+
+[tool.setuptools.packages.find]
+include = ["amdsmi*"]

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -32,7 +32,7 @@ COPY --link --from=rocm-source /opt/rocm /opt/rocm
 # ROCm environment variables - split to ensure ROCM_HOME is set first
 ENV ROCM_HOME=/opt/rocm
 ENV PATH="/opt/rocm/bin:/opt/rocm/llvm/bin:/home/ray/anaconda3/bin:/home/ray/.local/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin" \
-    LD_LIBRARY_PATH="/opt/rocm/lib:/opt/rocm/lib64" \
+    LD_LIBRARY_PATH="/opt/rocm/lib:/opt/rocm/lib64:/home/ray/anaconda3/lib" \
     HSA_PATH="/opt/rocm/hsa" \
     HIP_PATH="/opt/rocm/hip" \
     # Strix Halo (gfx1151) specific settings
@@ -84,6 +84,15 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
 RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
     uv pip install --system 'pandas>=2.0.0,<3.0'
 
+# Install amdsmi sysfs shim LAST (required for vLLM ROCm platform detection).
+# The native amdsmi from ROCm 7.1 requires glibc 2.38 (Ubuntu 24.04),
+# but the Ray base image is Ubuntu 22.04 (glibc 2.35). This pure-Python
+# shim reads GPU info from /sys/class/drm/* instead of libamd_smi.so.
+# Must be installed after vLLM/torch to prevent PyPI amdsmi from overwriting it.
+COPY --chown=1000:100 amdsmi-shim /tmp/amdsmi-shim
+RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
+    uv pip install --system /tmp/amdsmi-shim && rm -rf /tmp/amdsmi-shim
+
 # Pre-download common models for faster cold starts (optional, increases image size)
 # RUN python3 -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-large-en-v1.5')"