feat: full fan control — all fans or individual fans

The fan curve previously only controlled fan index 0; secondary fans
stayed on driver control. The curve can now target all fans (new
default) or any individual fan(s).

Backend:
- hal/fans.py: get_num_fans() via nvmlDeviceGetNumFans; get_fan_info()
  returns per-fan speeds; set_fan_speed() accepts a fan index list
  (None = all fans; all-fans mode is lenient toward driver-locked
  fans, explicit lists are strict); reset_fan() restores all fans.
- server.py: per-GPU fan_targets state; the poller applies the curve to
  all target fans and logs write failures (once per distinct error);
  activation validates targets against the hardware (stale indices fall
  back to all fans); POST /api/fans accepts fans, POST /api/fans/speed
  accepts a fan index, GET /api/fans returns num_fans/fans/fan_targets.
- Persistence format is now {"curve": ..., "fans": ...}; legacy
  bare-curve entries migrate to "all fans" at startup.
- Profiles save/apply fan_targets alongside fan_curve.
- MonitoringSample carries per-fan speeds for live gauges.

Frontend:
- Fans tab: All / Fan 1 / Fan 2 / ... selector with live per-fan %;
  the selection is applied together with the curve.
- Live monitor: per-fan gauges with sparklines for multi-fan GPUs.

nvcurve/config.py

@@ -34,8 +34,10 @@ class Config:
     # curve applied via the UI survives server restarts (fan control itself is
     # volatile — the driver reverts to automatic mode on reboot).
     # Key = stable GPU identifier (same as auto_load_profiles).
-    # Value = list of {"temp_c": int, "fan_pct": int} sorted by temp_c.
-    fan_curves: dict[str, list] = field(default_factory=dict)
+    # Value = {"curve": [{"temp_c": int, "fan_pct": int}, ...] sorted by temp_c,
+    #          "fans": [fan indices] | None (None = all fans)}.
+    # Legacy entries (bare curve list) are migrated at load time.
+    fan_curves: dict[str, object] = field(default_factory=dict)
 
 
 # Module-level default config instance.

nvcurve/hal/fans.py

@@ -1,10 +1,15 @@
 """Hardware Abstraction Layer for Fan Control.
 
 Uses NVML (via pynvml) for all operations:
-  - nvmlDeviceGetFanSpeed_v2   : read current fan speed % for a fan index
-  - nvmlDeviceSetFanSpeed_v2   : set fan speed % for a fan index
-  - nvmlDeviceGetMinMaxFanSpeed: get min/max fan speed constraints
-  - nvmlDeviceGetTemperature   : read GPU temp for curve interpolation
+  - nvmlDeviceGetNumFans             : number of fans on the device
+  - nvmlDeviceGetFanSpeed_v2         : read current fan speed % for a fan index
+  - nvmlDeviceSetFanSpeed_v2         : set fan speed % for a fan index
+  - nvmlDeviceSetDefaultFanSpeed_v2  : restore automatic control for a fan index
+  - nvmlDeviceGetMinMaxFanSpeed      : get min/max fan speed constraints
+  - nvmlDeviceGetTemperature         : read GPU temp for curve interpolation
+
+Fans are addressed by 0-based index. Passing ``fans=None`` to the set/reset
+helpers means "all fans on the device".
 """
 
 import ctypes
@@ -24,9 +29,6 @@ pynvml: Any = _pynvml_import
 
 log = logging.getLogger("nvcurve.hal.fans")
 
-# We use fan index 0 (first/primary fan) for all operations.
-_FAN_INDEX = 0
-
 
 def _get_handle(gpu_index: int):
     """Return an NVML device handle."""
@@ -35,13 +37,42 @@ def _get_handle(gpu_index: int):
     return pynvml.nvmlDeviceGetHandleByIndex(gpu_index)
 
 
+def _num_fans(handle) -> int:
+    """Return the number of fans on the device (>= 1 on query failure)."""
+    try:
+        return max(0, int(pynvml.nvmlDeviceGetNumFans(handle)))
+    except pynvml.NVMLError:
+        # GetNumFans unsupported: assume at least the primary fan exists.
+        return 1
+
+
+def get_num_fans(gpu_index: int = 0) -> int:
+    """Return the number of fans on the GPU (0 if NVML is unavailable)."""
+    if not _NVML_AVAILABLE:
+        return 0
+    try:
+        return _num_fans(_get_handle(gpu_index))
+    except pynvml.NVMLError as exc:
+        log.warning("get_num_fans: %s", exc)
+        return 0
+
+
 def get_fan_info(gpu_index: int = 0) -> dict:
-    """Return current fan state: fan_pct, fan_mode, min_fan_pct, max_fan_pct.
+    """Return current fan state for all fans.
+
+    Returns a dict with:
+      fan_pct     : current speed % of fan 0 (legacy, None on failure)
+      fans        : [{"index": i, "fan_pct": pct | None}, ...] per fan
+      num_fans    : number of fans on the device
+      fan_mode    : None (the server derives "auto"/"curve")
+      min_fan_pct / max_fan_pct : device-wide speed constraints
 
     Returns None values on failure.
     """
-    out: dict[str, float | None] = {
+    out: dict[str, Any] = {
         "fan_pct": None,
+        "fans": [],
+        "num_fans": 0,
         "fan_mode": None,
         "min_fan_pct": None,
         "max_fan_pct": None,
@@ -50,36 +81,54 @@ def get_fan_info(gpu_index: int = 0) -> dict:
         return out
     try:
         handle = _get_handle(gpu_index)
-
-        # Get current fan speed using v2 API (fan index 0)
-        try:
-            out["fan_pct"] = float(pynvml.nvmlDeviceGetFanSpeed_v2(handle, _FAN_INDEX))
-        except pynvml.NVMLError:
-            # Fallback to legacy v1 API
-            try:
-                out["fan_pct"] = float(pynvml.nvmlDeviceGetFanSpeed(handle))
-            except pynvml.NVMLError:
-                pass
-
-        # Get min/max fan speed constraints
-        try:
-            min_s = ctypes.c_uint(0)
-            max_s = ctypes.c_uint(0)
-            pynvml.nvmlDeviceGetMinMaxFanSpeed(handle, min_s, max_s)
-            out["min_fan_pct"] = int(min_s.value)
-            out["max_fan_pct"] = int(max_s.value)
-        except pynvml.NVMLError:
-            out["min_fan_pct"] = 0
-            out["max_fan_pct"] = 100
-
     except pynvml.NVMLError as exc:
         log.warning("get_fan_info: %s", exc)
+        return out
+
+    out["num_fans"] = _num_fans(handle)
+
+    # Per-fan speeds via the v2 API; fan 0 falls back to the legacy v1 API.
+    for i in range(out["num_fans"]):
+        pct: float | None = None
+        try:
+            pct = float(pynvml.nvmlDeviceGetFanSpeed_v2(handle, i))
+        except pynvml.NVMLError:
+            if i == 0:
+                try:
+                    pct = float(pynvml.nvmlDeviceGetFanSpeed(handle))
+                except pynvml.NVMLError:
+                    pct = None
+        out["fans"].append({"index": i, "fan_pct": pct})
+
+    out["fan_pct"] = out["fans"][0]["fan_pct"] if out["fans"] else None
+
+    # Get min/max fan speed constraints
+    try:
+        min_s = ctypes.c_uint(0)
+        max_s = ctypes.c_uint(0)
+        pynvml.nvmlDeviceGetMinMaxFanSpeed(handle, min_s, max_s)
+        out["min_fan_pct"] = int(min_s.value)
+        out["max_fan_pct"] = int(max_s.value)
+    except pynvml.NVMLError:
+        out["min_fan_pct"] = 0
+        out["max_fan_pct"] = 100
 
     return out
 
 
-def set_fan_speed(gpu_index: int, pct: int) -> tuple[bool, str]:
-    """Set fan speed to a percentage (0-100) on the primary fan."""
+def set_fan_speed(
+    gpu_index: int,
+    pct: int,
+    fans: list[int] | None = None,
+) -> tuple[bool, str]:
+    """Set fan speed to a percentage (0-100).
+
+    fans=None targets every fan on the device; fans=[0, 1] targets the
+    listed fan indices. In all-fans mode, fans the driver does not allow
+    manual control of (e.g. driver-mirrored secondary fans) are skipped
+    with a warning instead of failing the whole operation; explicit fan
+    lists are strict and fail if any selected fan cannot be set.
+    """
     try:
         pct = max(0, min(100, int(pct)))
     except (TypeError, ValueError):
@@ -88,7 +137,47 @@ def set_fan_speed(gpu_index: int, pct: int) -> tuple[bool, str]:
         return False, "NVML not available"
     try:
         handle = _get_handle(gpu_index)
-        pynvml.nvmlDeviceSetFanSpeed_v2(handle, _FAN_INDEX, pct)
+        num_fans = _num_fans(handle)
+
+        if fans is None:
+            targets = list(range(num_fans))
+            strict = False
+        else:
+            targets: list[int] = []
+            for f in fans:
+                try:
+                    f = int(f)
+                except (TypeError, ValueError):
+                    return False, f"Invalid fan index: {f!r}"
+                if f < 0 or f >= num_fans:
+                    return False, f"Fan index {f} out of range (0-{num_fans - 1})"
+                targets.append(f)
+            strict = True
+            if not targets:
+                return False, "No fans selected"
+
+        if not targets:
+            return False, "No fans available on this GPU"
+
+        skipped: list[str] = []
+        for i in targets:
+            try:
+                pynvml.nvmlDeviceSetFanSpeed_v2(handle, i, pct)
+            except pynvml.NVMLError as exc:
+                if strict:
+                    log.warning(
+                        "set_fan_speed(%d, fan %d, %d): %s", gpu_index, i, pct, exc
+                    )
+                    return False, str(exc)
+                # All-fans mode: secondary fans may be driver-controlled and
+                # reject manual writes; skip them and report in the message.
+                log.debug("set_fan_speed: fan %d not settable: %s", i, exc)
+                skipped.append(f"fan {i + 1}")
+        if skipped:
+            return (
+                True,
+                f"OK ({len(skipped)} fan(s) not manually controllable: {', '.join(skipped)})",
+            )
         return True, "OK"
     except pynvml.NVMLError as exc:
         log.warning("set_fan_speed(%d, %d): %s", gpu_index, pct, exc)
@@ -96,10 +185,11 @@ def set_fan_speed(gpu_index: int, pct: int) -> tuple[bool, str]:
 
 
 def reset_fan(gpu_index: int = 0) -> tuple[bool, str]:
-    """Restore automatic fan control.
+    """Restore automatic fan control for all fans.
 
-    Tries nvidia-smi --fan=default first (most reliable), then falls back to
-    NVML nvmlDeviceSetDefaultFanSpeed_v2.
+    Tries nvidia-smi --fan=default first (most reliable, resets all fans on
+    the device), then falls back to NVML nvmlDeviceSetDefaultFanSpeed_v2
+    per fan index.
     """
     if not _NVML_AVAILABLE:
         return False, "NVML not available"
@@ -124,10 +214,14 @@ def reset_fan(gpu_index: int = 0) -> tuple[bool, str]:
     except Exception as exc:
         log.debug("nvidia-smi -fan default error: %s", exc)
 
-    # Fallback: use NVML to reset to default fan speed
+    # Fallback: use NVML to reset every fan to default speed
     try:
         handle = _get_handle(gpu_index)
-        pynvml.nvmlDeviceSetDefaultFanSpeed_v2(handle, _FAN_INDEX)
+        for i in range(_num_fans(handle)):
+            try:
+                pynvml.nvmlDeviceSetDefaultFanSpeed_v2(handle, i)
+            except pynvml.NVMLError as exc:
+                log.debug("reset_fan: fan %d: %s", i, exc)
         return True, "OK"
     except pynvml.NVMLError as exc:
         return False, f"Failed to reset fan: {exc}"

nvcurve/hal/monitoring.py

@@ -162,6 +162,20 @@ def _nvml_read(gpu_index: int) -> dict:
         with contextlib.suppress(_pynvml.NVMLError):
             out["fan_pct"] = float(_pynvml.nvmlDeviceGetFanSpeed(handle))
 
+        # Per-fan speeds via the v2 API (fan_pct above stays fan 0 for legacy clients).
+        try:
+            num_fans = int(_pynvml.nvmlDeviceGetNumFans(handle))
+            fan_list: list[float | None] = []
+            for i in range(num_fans):
+                try:
+                    fan_list.append(float(_pynvml.nvmlDeviceGetFanSpeed_v2(handle, i)))
+                except _pynvml.NVMLError:
+                    fan_list.append(None)
+            if fan_list:
+                out["fans"] = fan_list
+        except _pynvml.NVMLError:
+            pass
+
         with contextlib.suppress(_pynvml.NVMLError):
             out["throttle_reasons"] = int(
                 _pynvml.nvmlDeviceGetCurrentClocksThrottleReasons(handle)

nvcurve/nvapi/types.py

@@ -61,6 +61,7 @@ class MonitoringSample:
     pcie_link_width: int | None = None  # Current PCIe link width (x1..x16)
     pcie_link_generation: int | None = None  # Current PCIe link generation (1..5)
     mem_temp_c: float | None = None  # VRAM temperature (if the GPU exposes it)
+    fans: list[float | None] | None = None  # Per-fan speed % (index-aligned)
 
 
 @dataclass

nvcurve/profiles/native.py

@@ -17,6 +17,8 @@ class ProfileData:
     mem_offset_mhz: int | None = None
     power_limit_w: int | None = None
     fan_curve: list[dict[str, int]] | None = None
+    # Fan indices controlled by fan_curve (0-based); None = all fans.
+    fan_targets: list[int] | None = None
 
 
 def save_profile(profile_dir: str, data: ProfileData) -> str:

nvcurve/server.py

@@ -24,6 +24,7 @@ from .config import Config, default_config
 from .hal.dashboard import get_dashboard_info
 from .hal.fans import (
     get_fan_info,
+    get_num_fans,
     get_temp,
     interpolate_fan_speed,
     reset_fan,
@@ -149,6 +150,7 @@ def _sample_dict(s) -> dict:
         "temp_c": s.temp_c,
         "power_w": s.power_w,
         "fan_pct": s.fan_pct,
+        "fans": s.fans,
         "pstate": s.pstate,
         "pstate_label": f"P{s.pstate}" if s.pstate is not None else None,
         "mem_used_bytes": s.mem_used_bytes,
@@ -219,7 +221,8 @@ async def _monitor_poller(gpu_index: int) -> None:
 
 
 async def _fan_poller(gpu_index: int) -> None:
-    """Continuously read GPU temp, interpolate fan speed from active curve, and apply."""
+    """Continuously read GPU temp, interpolate fan speed from active curve, and apply it to the target fans."""
+    last_write_error: str | None = None
     while True:
         try:
             g_state = _state["gpus"].get(gpu_index)
@@ -229,7 +232,25 @@ async def _fan_poller(gpu_index: int) -> None:
                     curve = g_state["fan_curve"]
                     target = interpolate_fan_speed(curve, temp)
                     if target is not None:
-                        await _run(set_fan_speed, gpu_index, target)
+                        ok, msg = await _run(
+                            set_fan_speed,
+                            gpu_index,
+                            target,
+                            g_state.get("fan_targets"),
+                        )
+                        if not ok:
+                            # Log once per distinct failure so a stuck target
+                            # list doesn't spam a warning every 2 s tick.
+                            if msg != last_write_error:
+                                log.warning(
+                                    "Fan write failed for GPU %d (targets=%s): %s",
+                                    gpu_index,
+                                    g_state.get("fan_targets"),
+                                    msg,
+                                )
+                            last_write_error = msg
+                        else:
+                            last_write_error = None
         except asyncio.CancelledError:
             return
         except Exception as exc:
@@ -237,15 +258,34 @@ async def _fan_poller(gpu_index: int) -> None:
         await asyncio.sleep(2.0)
 
 
-async def _activate_fan_curve(gpu_index: int, curve: list) -> None:
+async def _activate_fan_curve(
+    gpu_index: int, curve: list, fans: list[int] | None = None
+) -> None:
     """Set the active fan curve, (re)start the poller, and persist it.
 
-    Persistence (config.json) is what makes the curve survive server restarts:
-    fan control is volatile, so the driver reverts to automatic mode on reboot
-    and the saved curve is re-applied at the next server start.
+    fans=None targets all fans on the device; a list targets the given
+    fan indices. Persistence (config.json) is what makes the curve survive
+    server restarts: fan control is volatile, so the driver reverts to
+    automatic mode on reboot and the saved curve is re-applied at the next
+    server start.
     """
     g_state = _get_gpu_state(gpu_index)
 
+    # Validate explicit fan targets against the hardware. Stale indices
+    # (e.g. a profile saved on a 2-fan GPU applied to a 1-fan GPU, or a
+    # persisted entry restored after a hardware change) would otherwise
+    # make the poller fail silently on every tick.
+    if fans is not None:
+        num_fans = await _run(get_num_fans, gpu_index)
+        if not fans or (num_fans > 0 and any(f < 0 or f >= num_fans for f in fans)):
+            log.warning(
+                "Fan targets %s invalid for GPU %d (%d fan(s)); falling back to all fans",
+                fans,
+                gpu_index,
+                num_fans,
+            )
+            fans = None
+
     # Stop existing poller if running
     if g_state.get("fan_poller_task"):
         g_state["fan_poller_task"].cancel()
@@ -254,10 +294,11 @@ async def _activate_fan_curve(gpu_index: int, curve: list) -> None:
 
     g_state["fan_curve"] = curve
     g_state["fan_curve_active"] = True
+    g_state["fan_targets"] = fans
     g_state["fan_poller_task"] = asyncio.create_task(_fan_poller(gpu_index))
 
     cfg: Config = _state["config"]
-    cfg.fan_curves[_gpu_stable_key(gpu_index)] = curve
+    cfg.fan_curves[_gpu_stable_key(gpu_index)] = {"curve": curve, "fans": fans}
     _persist_fan_curves(cfg.fan_curves)
 
 
@@ -276,6 +317,7 @@ async def _deactivate_fan_curve(gpu_index: int, reset_hardware: bool = True) ->
 
     g_state["fan_curve"] = None
     g_state["fan_curve_active"] = False
+    g_state["fan_targets"] = None
 
     if reset_hardware:
         ok, msg = await _run(reset_fan, gpu_index)
@@ -389,7 +431,15 @@ async def lifespan(app: FastAPI):
         if g_state.get("fan_curve_active"):
             continue  # already activated by the auto-load profile path
         key = _gpu_stable_key(gpu_index)
-        curve = cfg.fan_curves.get(key)
+        entry = cfg.fan_curves.get(key)
+        if not entry:
+            continue
+        # Migrate the legacy format (bare curve list) to the current
+        # {"curve": ..., "fans": ...} shape; legacy entries targeted all fans.
+        if isinstance(entry, list):
+            entry = {"curve": entry, "fans": None}
+        curve = entry.get("curve") if isinstance(entry, dict) else None
+        fans = entry.get("fans") if isinstance(entry, dict) else None
         if not curve:
             continue
         ok, msg = validate_curve(curve)
@@ -398,7 +448,7 @@ async def lifespan(app: FastAPI):
             continue
         log.info("Restoring persisted fan curve on GPU %d (%s)", gpu_index, key)
         try:
-            await _activate_fan_curve(gpu_index, curve)
+            await _activate_fan_curve(gpu_index, curve, fans)
         except Exception as exc:
             log.warning(
                 "Failed to restore persisted fan curve on GPU %d: %s",
@@ -543,10 +593,14 @@ class FanCurvePoint(BaseModel):
 
 class FanCurveRequest(BaseModel):
     curve: list[FanCurvePoint]
+    # Fan indices to control (0-based); None = all fans on the device.
+    fans: list[int] | None = None
 
 
 class FanSpeedRequest(BaseModel):
     fan_pct: int
+    # Specific fan index to set (0-based); None = all fans on the device.
+    fan: int | None = None
 
 
 class LoginRequest(BaseModel):
@@ -873,6 +927,9 @@ async def api_profile_save(req: ProfileSaveRequest, gpu_index: int = 0):
         mem_offset_mhz=mem_offset_mhz,
         power_limit_w=power_limit_w,
         fan_curve=g_state.get("fan_curve") if g_state.get("fan_curve_active") else None,
+        fan_targets=g_state.get("fan_targets")
+        if g_state.get("fan_curve_active")
+        else None,
     )
     filepath = await _run(save_profile, cfg.profile_dir, data)
     g_state["active_profile"] = req.name
@@ -1024,7 +1081,7 @@ async def _apply_profile(name: str, gpu_index: int = 0) -> list[str]:
         if not ok:
             errs.append(f"Fan curve: {msg}")
         else:
-            await _activate_fan_curve(gpu_index, profile.fan_curve)
+            await _activate_fan_curve(gpu_index, profile.fan_curve, profile.fan_targets)
     elif g_state.get("fan_curve_active"):
         await _deactivate_fan_curve(gpu_index, reset_hardware=True)
 
@@ -1226,7 +1283,7 @@ async def api_limits_reset(gpu_index: int = 0):
 
 @app.get("/api/fans")
 async def api_fans(gpu_index: int = 0):
-    """Current fan state: fan %, curve, and whether curve control is active."""
+    """Current fan state: per-fan %, curve, and whether curve control is active."""
     _get_gpu_state(gpu_index)
     g_state = _state["gpus"][gpu_index]
     info = await _run(get_fan_info, gpu_index)
@@ -1236,12 +1293,16 @@ async def api_fans(gpu_index: int = 0):
         "fan_mode": "curve" if curve_active else "auto",
         "curve": g_state.get("fan_curve"),
         "curve_active": curve_active,
+        "fan_targets": g_state.get("fan_targets"),
     }
 
 
 @app.post("/api/fans")
 async def api_fans_update(req: FanCurveRequest, gpu_index: int = 0):
-    """Set or update the fan curve. Starts the fan control poller."""
+    """Set or update the fan curve. Starts the fan control poller.
+
+    req.fans selects which fans the curve drives (None = all fans).
+    """
     g_state = _get_gpu_state(gpu_index)
 
     curve_data = [{"temp_c": p.temp_c, "fan_pct": p.fan_pct} for p in req.curve]
@@ -1249,6 +1310,8 @@ async def api_fans_update(req: FanCurveRequest, gpu_index: int = 0):
     if not ok:
         raise HTTPException(status_code=400, detail=msg)
 
+    fans = sorted(set(req.fans)) if req.fans is not None else None
+
     # Test that fan control is available on this GPU, probing with the
     # target for the *current* temperature so the fan is never briefly
     # set to an inappropriate speed.
@@ -1261,14 +1324,14 @@ async def api_fans_update(req: FanCurveRequest, gpu_index: int = 0):
         )
         target = interpolate_fan_speed(curve_data, test_temp)
         if target is not None:
-            fan_ok, fan_msg = await _run(set_fan_speed, gpu_index, target)
+            fan_ok, fan_msg = await _run(set_fan_speed, gpu_index, target, fans)
             if not fan_ok:
                 raise HTTPException(
                     status_code=500, detail=f"Fan control not available: {fan_msg}"
                 )
 
     # Activate the curve (starts the poller) and persist it so it survives restarts.
-    await _activate_fan_curve(gpu_index, curve_data)
+    await _activate_fan_curve(gpu_index, curve_data, fans)
 
     return {"ok": True}
 
@@ -1287,10 +1350,14 @@ async def api_fans_reset(gpu_index: int = 0):
 
 @app.post("/api/fans/speed")
 async def api_fans_speed(req: FanSpeedRequest, gpu_index: int = 0):
-    """One-shot set fan to an exact percentage (bypasses curve)."""
+    """One-shot set fan(s) to an exact percentage (bypasses curve).
+
+    req.fan selects a single fan index; None sets all fans.
+    """
     _get_gpu_state(gpu_index)
     pct = max(0, min(100, req.fan_pct))
-    ok, msg = await _run(set_fan_speed, gpu_index, pct)
+    fans = [req.fan] if req.fan is not None else None
+    ok, msg = await _run(set_fan_speed, gpu_index, pct, fans)
     if not ok:
         raise HTTPException(status_code=500, detail=msg)
     return {"ok": True}