"""VRAM Arbitrator and High-Speed Switch Manager for Ollama and ComfyUI.""" import time import httpx import psutil import logging from typing import Dict, List, Any, Optional from collections import deque import asyncio import json import websockets import overclock_manager import ram_optimizer import telemetry_store try: import pynvml pynvml.nvmlInit() NVML_AVAILABLE = True except Exception as e: NVML_AVAILABLE = False logger = logging.getLogger("vram_arbitrator") OLLAMA_API_BASE = "http://localhost:11434" COMFY_API_BASE = "http://127.0.0.1:8188" # Circular buffer for transition events (the durable log lives in telemetry_store) SWITCH_HISTORY = deque(maxlen=50) # Bandwidth thresholds for classifying how a model reached VRAM, calibrated by measuring # the same 12.87 GB model loaded cold and warm on this box (2026-08-28): # # 3.1% resident -> 34.3 s -> 0.38 GB/s # 100% resident -> 4.9 s -> 2.63 GB/s # # The first cut at these numbers assumed a page-cache-fed load would approach the bus # rate and set the cache-hit bar at 5 GB/s. It does not: Ollama's load_duration covers # host-to-device transfer and model initialisation as well as the file read, so a fully # resident model still reports ~2.6 GB/s while the page cache itself reads at 6.4 GB/s. # A 5 GB/s bar could therefore never be met, and every warm load was being reported as # a partial hit. Thresholds now sit either side of the measured 6.9x separation. RAM_HIT_GBPS = 2.0 PARTIAL_HIT_GBPS = 0.8 # How long to wait for Ollama's VRAM to actually drain. # # Ollama will not unload a model mid-generation. With OLLAMA_NUM_PARALLEL=1 our # keep_alive:0 request queues behind the running one and takes effect the moment it # finishes, so a model that is busy is not failing -- it is finishing, and it will # release on its own. Blocking the arbitrator for the length of someone's inference # helps nobody: ComfyUI is not gated on our return value, and every second spent # blocked is a second the watchdog and profile switching are stalled. # # So: wait briefly for the common case (an idle model releases in 40-110 ms here), # then classify. A caller who genuinely wants to block can ask for a longer wait. YIELD_CONFIRM_TIMEOUT_S = 2.0 YIELD_CONFIRM_TIMEOUT_BLOCKING_S = 30.0 # A model still holding VRAM while the GPU is pinned is generating, not wedged. BUSY_UTIL_PCT = 50 BUSY_PROBE_S = 0.6 # Fraction of a model that may sit outside VRAM before we call it starved. A little # slack absorbs rounding and KV-cache accounting; beyond it, layers are on the CPU. CPU_OFFLOAD_TOLERANCE = 0.02 # Only intervene when ComfyUI is actually holding enough VRAM to be the cause. RECLAIM_MIN_COMFY_BYTES = 512 * 1024 ** 2 # Ollama's response when a model will not fit. Which of the two failure modes you get # depends on configuration: with n_gpu_layers left to Ollama it spills layers to the CPU # and reports size_vram < size; with n_gpu_layers pinned (99 on this box) it refuses and # returns a hard CUDA OOM instead. Both are handled -- the spill by # AutoArbitrator._check_ollama_starved, the hard failure by the retry below. OOM_SIGNATURES = ("out of memory", "cudamalloc", "unable to allocate", "failed to allocate", "cuda error") def looks_like_vram_oom(text: str) -> bool: low = (text or "").lower() return any(sig in low for sig in OOM_SIGNATURES) YIELD_CONFIRM_POLL_S = 0.02 YIELD_RESIDUAL_BYTES = 256 * 1024 ** 2 # treat <256 MB as "released" # Connection-pooled clients. Re-creating an AsyncClient per call meant a fresh TCP # handshake on every one of the watchdog's polls. _clients: Dict[str, httpx.AsyncClient] = {} def _client(base_url: str, timeout: float) -> httpx.AsyncClient: key = f"{base_url}|{timeout}" c = _clients.get(key) if c is None or c.is_closed: c = httpx.AsyncClient( base_url=base_url, timeout=timeout, limits=httpx.Limits(max_keepalive_connections=4, max_connections=8), ) _clients[key] = c return c async def close_clients() -> None: for c in list(_clients.values()): try: await c.aclose() except Exception: pass _clients.clear() # Bit flags from nvmlDeviceGetCurrentClocksThrottleReasons, decoded for the governor. THROTTLE_REASONS = { 0x0000000000000001: "gpu_idle", 0x0000000000000002: "applications_clocks_setting", 0x0000000000000004: "sw_power_cap", 0x0000000000000008: "hw_slowdown", 0x0000000000000010: "sync_boost", 0x0000000000000020: "sw_thermal_slowdown", 0x0000000000000040: "hw_thermal_slowdown", 0x0000000000000080: "hw_power_brake_slowdown", 0x0000000000000100: "display_clock_setting", } def decode_throttle_reasons(bits: int) -> List[str]: return [name for mask, name in THROTTLE_REASONS.items() if bits & mask] def get_process_vram_bytes() -> Dict[str, int]: """Fast NVML-only VRAM attribution, used by the yield barrier's tight poll loop. Deliberately avoids psutil lookups: this runs every 20 ms while we wait for VRAM to actually drain. """ out = {"ollama_bytes": 0, "comfyui_bytes": 0, "other_bytes": 0, "free_bytes": 0, "gpu_util_pct": 0} if not NVML_AVAILABLE: return out try: handle = pynvml.nvmlDeviceGetHandleByIndex(0) out["free_bytes"] = pynvml.nvmlDeviceGetMemoryInfo(handle).free try: out["gpu_util_pct"] = pynvml.nvmlDeviceGetUtilizationRates(handle).gpu except Exception: pass procs = list(pynvml.nvmlDeviceGetComputeRunningProcesses(handle)) try: procs += list(pynvml.nvmlDeviceGetGraphicsRunningProcesses(handle)) except Exception: pass merged: Dict[int, int] = {} for p in procs: merged[p.pid] = max(merged.get(p.pid, 0), p.usedGpuMemory or 0) for pid, used in merged.items(): kind = _PID_KIND_CACHE.get(pid) if kind is None: kind = _classify_pid(pid) _PID_KIND_CACHE[pid] = kind if kind == "ollama": out["ollama_bytes"] += used elif kind == "comfy": out["comfyui_bytes"] += used else: out["other_bytes"] += used except Exception as e: logger.debug(f"get_process_vram_bytes failed: {e}") return out _PID_KIND_CACHE: Dict[int, str] = {} def _classify_pid(pid: int) -> str: try: proc = psutil.Process(pid) pname = proc.name().lower() cmdline = " ".join(proc.cmdline()).lower() except Exception: return "other" if "ollama" in pname or "llama-server" in cmdline: return "ollama" if "comfy" in cmdline or "main.py" in cmdline: return "comfy" return "other" def get_gpu_hardware_stats() -> Dict[str, Any]: """Retrieve comprehensive GPU hardware and process metrics via NVML.""" if not NVML_AVAILABLE: return {"available": False, "error": "NVML not initialized"} try: handle = pynvml.nvmlDeviceGetHandleByIndex(0) name = pynvml.nvmlDeviceGetName(handle) if isinstance(name, bytes): name = name.decode("utf-8") mem_info = pynvml.nvmlDeviceGetMemoryInfo(handle) util_rates = pynvml.nvmlDeviceGetUtilizationRates(handle) temp_c = pynvml.nvmlDeviceGetTemperature(handle, pynvml.NVML_TEMPERATURE_GPU) try: power_mw = pynvml.nvmlDeviceGetPowerUsage(handle) power_w = round(power_mw / 1000.0, 1) except Exception: power_w = 0.0 fan_pct = 0 fans = [] try: num_fans = pynvml.nvmlDeviceGetNumFans(handle) for i in range(num_fans): try: fans.append(pynvml.nvmlDeviceGetFanSpeed_v2(handle, i)) except Exception: pass if fans: fan_pct = max(fans) else: fan_pct = pynvml.nvmlDeviceGetFanSpeed(handle) fans = [fan_pct] except Exception: try: fan_pct = pynvml.nvmlDeviceGetFanSpeed(handle) fans = [fan_pct] except Exception: fan_pct = 0 fans = [] try: clock_graphics = pynvml.nvmlDeviceGetClockInfo(handle, pynvml.NVML_CLOCK_GRAPHICS) clock_mem = pynvml.nvmlDeviceGetClockInfo(handle, pynvml.NVML_CLOCK_MEM) except Exception: clock_graphics = 0 clock_mem = 0 # PCIe throughput (KB/s) — TX + RX. Key metric for the RAM-cache # PCIe-speed swap thesis (assimilated from pmady/gpu-mcp-server). pcie_tx_kbps = 0 pcie_rx_kbps = 0 try: pcie_tx_kbps = pynvml.nvmlDeviceGetPcieThroughput(handle, pynvml.NVML_PCIE_UTIL_TX_BYTES) except Exception: pass try: pcie_rx_kbps = pynvml.nvmlDeviceGetPcieThroughput(handle, pynvml.NVML_PCIE_UTIL_RX_BYTES) except Exception: pass # Why the GPU is not running at full clocks — the thermal governor reads this. throttle_bits = 0 throttle_reasons: List[str] = [] try: throttle_bits = pynvml.nvmlDeviceGetCurrentClocksThrottleReasons(handle) throttle_reasons = decode_throttle_reasons(throttle_bits) except Exception: pass # Power management limit (watts) — the OC ceiling. power_limit_w = 0.0 try: power_limit_w = round(pynvml.nvmlDeviceGetPowerManagementLimit(handle) / 1000.0, 1) except Exception: pass # Driver + CUDA version (completeness). driver_version = "" cuda_version = "" try: dv = pynvml.nvmlSystemGetDriverVersion() driver_version = dv.decode("utf-8") if isinstance(dv, bytes) else str(dv) except Exception: pass try: cv = pynvml.nvmlSystemGetCudaDriverVersion() cuda_version = cv # int like 12030 == CUDA 12.3 except Exception: pass # Discover processes on GPU proc_breakdown = { "ollama_bytes": 0, "comfyui_bytes": 0, "system_bytes": 0, "processes": [] } try: procs = pynvml.nvmlDeviceGetComputeRunningProcesses(handle) graphics_procs = pynvml.nvmlDeviceGetGraphicsRunningProcesses(handle) all_procs = {p.pid: p.usedGpuMemory for p in procs} for p in graphics_procs: all_procs[p.pid] = max(all_procs.get(p.pid, 0), p.usedGpuMemory or 0) for pid, used_mem in all_procs.items(): pname = "Unknown" cmdline = "" try: proc = psutil.Process(pid) pname = proc.name() cmdline = " ".join(proc.cmdline()) except Exception: pass is_ollama = "ollama" in pname.lower() or "llama-server" in cmdline.lower() is_comfy = "comfy" in cmdline.lower() or "main.py" in cmdline.lower() if is_ollama: proc_breakdown["ollama_bytes"] += used_mem elif is_comfy: proc_breakdown["comfyui_bytes"] += used_mem else: proc_breakdown["system_bytes"] += used_mem proc_breakdown["processes"].append({ "pid": pid, "name": pname, "cmdline": cmdline[:60], "vram_bytes": used_mem, "vram_mb": round(used_mem / (1024**2), 1), "is_ollama": is_ollama, "is_comfy": is_comfy, }) except Exception as e: logger.error(f"Error enumerating GPU processes: {e}") total_vram = mem_info.total used_vram = mem_info.used free_vram = mem_info.free return { "available": True, "device_name": name, "vram_total_bytes": total_vram, "vram_total_gb": round(total_vram / (1024**3), 2), "vram_used_bytes": used_vram, "vram_used_gb": round(used_vram / (1024**3), 2), "vram_free_bytes": free_vram, "vram_free_gb": round(free_vram / (1024**3), 2), "vram_used_pct": round((used_vram / total_vram * 100) if total_vram > 0 else 0, 1), "gpu_util_pct": util_rates.gpu, "mem_util_pct": util_rates.memory, "temperature_c": temp_c, "power_w": power_w, "power_limit_w": power_limit_w, "pcie_tx_kbps": pcie_tx_kbps, "pcie_rx_kbps": pcie_rx_kbps, "throttle_bits": throttle_bits, "throttle_reasons": throttle_reasons, "driver_version": driver_version, "cuda_version": cuda_version, "fan_pct": fan_pct, "fans": fans, "num_fans": len(fans), "clock_graphics_mhz": clock_graphics, "clock_mem_mhz": clock_mem, "breakdown": { "ollama_mb": round(proc_breakdown["ollama_bytes"] / (1024**2), 1), "ollama_gb": round(proc_breakdown["ollama_bytes"] / (1024**3), 2), "comfyui_mb": round(proc_breakdown["comfyui_bytes"] / (1024**2), 1), "comfyui_gb": round(proc_breakdown["comfyui_bytes"] / (1024**3), 2), "system_mb": round(proc_breakdown["system_bytes"] / (1024**2), 1), "system_gb": round(proc_breakdown["system_bytes"] / (1024**3), 2), "free_mb": round(free_vram / (1024**2), 1), "free_gb": round(free_vram / (1024**3), 2), "processes": proc_breakdown["processes"], } } except Exception as e: return {"available": False, "error": str(e)} async def get_ollama_live_state() -> Dict[str, Any]: """Get active models, running status, and VRAM expiration from Ollama.""" state = { "online": False, "loaded_models": [], "active_model_name": None, "active_model_vram_gb": 0.0, "active_context": 0, "expires_at": None, "installed_models": [], # Ollama silently spills layers to CPU when VRAM is short. size_vram < size is the # only externally visible sign, and the cost is roughly an order of magnitude in # decode speed, so it is worth surfacing loudly. "gpu_fraction": 1.0, "cpu_offload_pct": 0.0, "partially_offloaded": False, } try: client = _client(OLLAMA_API_BASE, 3.0) # Check running models (ps) ps_resp = await client.get("/api/ps") if ps_resp.status_code == 200: state["online"] = True models = ps_resp.json().get("models", []) state["loaded_models"] = models if models: first = models[0] state["active_model_name"] = first.get("name") vram_bytes = first.get("size_vram", first.get("size", 0)) state["active_model_vram_gb"] = round(vram_bytes / (1024**3), 2) state["active_context"] = first.get("context_length", 0) state["expires_at"] = first.get("expires_at") # Check all tags tags_resp = await client.get("/api/tags") if tags_resp.status_code == 200: state["installed_models"] = tags_resp.json().get("models", []) except Exception as e: logger.debug(f"Ollama check error: {e}") return state async def get_comfyui_live_state() -> Dict[str, Any]: """Get prompt queue, device status, and active execution from ComfyUI.""" state = { "online": False, "executing": False, "queue_remaining": 0, "queue_running": 0, "current_node": None, "current_prompt_id": None, "vram_free_mb": 0, "vram_total_mb": 0, } try: client = _client(COMFY_API_BASE, 3.0) # Check system stats stats_resp = await client.get("/system_stats") if stats_resp.status_code == 200: state["online"] = True data = stats_resp.json() devices = data.get("devices", []) if devices: dev = devices[0] state["vram_free_mb"] = round(dev.get("vram_free", 0) / (1024**2), 1) state["vram_total_mb"] = round(dev.get("vram_total", 0) / (1024**2), 1) # Check queue queue_resp = await client.get("/queue") if queue_resp.status_code == 200: qdata = queue_resp.json() running = qdata.get("queue_running", []) pending = qdata.get("queue_pending", []) state["queue_running"] = len(running) state["queue_remaining"] = len(pending) state["executing"] = len(running) > 0 if running: state["current_prompt_id"] = running[0][1] if len(running[0]) > 1 else str(running[0]) except Exception as e: logger.debug(f"ComfyUI check error: {e}") return state async def _await_vram_release(baseline_bytes: int, timeout_s: float = YIELD_CONFIRM_TIMEOUT_S) -> Dict[str, Any]: """Wait for Ollama's VRAM to drain, distinguishing "busy" from "stuck". Posting keep_alive:0 only *asks* Ollama to unload; the driver frees the allocation some milliseconds later, and returning before that happens is how ComfyUI ends up allocating into VRAM that is still occupied. But there is a second case the first version of this got wrong. If the model is mid-generation it cannot unload at all, and reporting that as a timeout made a perfectly healthy cron job look like a 95% failure rate. When the VRAM has not moved and the GPU is pinned, the model is working; the queued unload will fire when it finishes. That is `busy`, not a failure. Returns an `outcome` of "released", "busy" or "stuck". """ t0 = time.perf_counter() peak_util = 0 while True: snap = get_process_vram_bytes() last = snap["ollama_bytes"] peak_util = max(peak_util, snap.get("gpu_util_pct", 0)) elapsed = time.perf_counter() - t0 if last <= YIELD_RESIDUAL_BYTES: return { "outcome": "released", "confirmed": True, "confirm_ms": round(elapsed * 1000, 2), "residual_bytes": last, "free_bytes": snap["free_bytes"], "gpu_util_pct": snap.get("gpu_util_pct", 0), } # Unmoved VRAM plus a pinned GPU means a generation is in flight. busy = (elapsed >= BUSY_PROBE_S and last >= baseline_bytes - YIELD_RESIDUAL_BYTES and peak_util >= BUSY_UTIL_PCT) if busy or elapsed >= timeout_s: outcome = "busy" if busy else "stuck" return { "outcome": outcome, "confirmed": False, "confirm_ms": round(elapsed * 1000, 2), "residual_bytes": last, "free_bytes": snap["free_bytes"], "gpu_util_pct": snap.get("gpu_util_pct", 0), "peak_util_pct": peak_util, "error": ( f"Ollama is mid-generation ({peak_util}% GPU, " f"{round(last / (1024**3), 2)} GB held); the queued unload will apply " f"when it finishes" if outcome == "busy" else f"Ollama still holding {round(last / (1024**3), 2)} GB after " f"{timeout_s}s with the GPU idle" ), } await asyncio.sleep(YIELD_CONFIRM_POLL_S) # Detached tasks need a strong reference or the loop may garbage-collect them mid-flight. _DETACHED: set = set() def _spawn_detached(coro) -> None: task = asyncio.ensure_future(coro) _DETACHED.add(task) task.add_done_callback(_DETACHED.discard) async def _confirm_release_later(targets: List[str], baseline_bytes: int, max_wait_s: float = 900.0) -> None: """Watch for a queued unload to land after the in-flight generation finishes. Runs detached so the caller is never held for the length of an inference. Logs the eventual release so the event log tells the whole story rather than stopping at "deferred". """ t0 = time.perf_counter() while (time.perf_counter() - t0) < max_wait_s: await asyncio.sleep(0.5) snap = get_process_vram_bytes() if snap["ollama_bytes"] <= YIELD_RESIDUAL_BYTES: waited_ms = round((time.perf_counter() - t0) * 1000, 2) _record({ "event_type": "Ollama VRAM Yield", "source": ", ".join(targets)[:200], "target": "VRAM 0MB (Kept in RAM)", "duration_ms": waited_ms, "yield_confirm_ms": waited_ms, "cache_status": "RAM-Cached", "detail": "released after the in-flight generation completed", }) arbitrator.note_deferred_release(waited_ms) logger.info(f"Deferred VRAM yield completed after {round(waited_ms / 1000, 1)}s " f"({round(snap['free_bytes'] / (1024**3), 2)} GB free)") return logger.warning("Deferred VRAM yield never landed within " f"{max_wait_s}s for {', '.join(targets)}") def _record(event: Dict[str, Any]) -> None: """Push an event to both the in-memory ring and the durable store.""" event.setdefault("ts", time.time()) event.setdefault("timestamp", time.strftime("%H:%M:%S")) SWITCH_HISTORY.appendleft(event) telemetry_store.record_event(event, profile=overclock_manager.ACTIVE_PROFILE) async def instant_free_ollama_vram(model_name: Optional[str] = None, confirm: bool = True, timeout_s: Optional[float] = None) -> Dict[str, Any]: """Yield Ollama's VRAM and wait for the driver to actually release it. The returned duration_ms is now the real end-to-end release time, not just how long the HTTP POST took. """ t0 = time.perf_counter() if model_name: targets = [model_name] else: # Unload *every* resident model, not just loaded_models[0]. Ollama will happily # keep several models in VRAM at once; releasing only the first left the rest # allocated, which the confirm barrier caught as "still holding 8.2 GB after 3s". ollama_state = await get_ollama_live_state() targets = [m.get("name") for m in ollama_state.get("loaded_models", []) if m.get("name")] if not targets and ollama_state.get("active_model_name"): targets = [ollama_state["active_model_name"]] if not targets: return {"success": True, "message": "No active Ollama model in VRAM", "duration_ms": 0, "confirmed": True} model_name = targets[0] if len(targets) == 1 else f"{len(targets)} models" baseline = get_process_vram_bytes()["ollama_bytes"] try: # Generous client timeout: with OLLAMA_NUM_PARALLEL=1 this request queues behind # any running generation, and a short timeout would drop the connection before # Ollama ever processed the unload -- losing it entirely. client = _client(OLLAMA_API_BASE, 120.0) unload_calls = [ client.post("/api/generate", json={"model": t, "keep_alive": 0}) for t in targets ] # Do not await the queued unloads; a busy model would block us for the length of # its inference. They are fire-and-confirm: the barrier below watches the VRAM. _spawn_detached(asyncio.gather(*unload_calls, return_exceptions=True)) request_ms = round((time.perf_counter() - t0) * 1000, 2) barrier: Dict[str, Any] = {"outcome": "unconfirmed", "confirmed": None, "confirm_ms": 0.0, "residual_bytes": baseline} if confirm: barrier = await _await_vram_release( baseline, timeout_s if timeout_s is not None else YIELD_CONFIRM_TIMEOUT_S) if barrier.get("outcome") == "busy": # The unload is queued and will fire when the generation ends. Keep watching # in the background so the release is still logged and the counters stay true, # without holding the caller here for the length of someone's inference. _spawn_detached(_confirm_release_later(targets, baseline)) duration_ms = round((time.perf_counter() - t0) * 1000, 2) freed_gb = round(max(baseline - barrier.get("residual_bytes", 0), 0) / (1024**3), 2) outcome = barrier.get("outcome", "unconfirmed") _record({ "event_type": "Ollama VRAM Yield", "source": ", ".join(targets)[:200], "target": "VRAM 0MB (Kept in RAM)", "duration_ms": duration_ms, "yield_confirm_ms": barrier.get("confirm_ms"), "cache_status": { "released": "RAM-Cached", "busy": "Deferred — LLM generating", "stuck": "Yield Stalled", }.get(outcome, "Yield Unconfirmed"), "detail": barrier.get("error"), }) return { "success": True, "outcome": outcome, "model": model_name, "models_unloaded": targets, "duration_ms": duration_ms, "request_ms": request_ms, "confirm_ms": barrier.get("confirm_ms"), "confirmed": barrier.get("confirmed"), "freed_gb": freed_gb, "residual_gb": round(barrier.get("residual_bytes", 0) / (1024**3), 2), "free_vram_gb": round(barrier.get("free_bytes", 0) / (1024**3), 2), "gpu_util_pct": barrier.get("gpu_util_pct"), "error": barrier.get("error"), } except Exception as e: return {"success": False, "error": str(e), "duration_ms": round((time.perf_counter() - t0) * 1000, 2)} async def instant_free_comfyui_vram() -> Dict[str, Any]: """Tell ComfyUI to purge loaded diffusion models from VRAM.""" t0 = time.perf_counter() try: client = _client(COMFY_API_BASE, 5.0) await client.post("/free", json={"unload_models": True, "free_memory": True}) duration_ms = round((time.perf_counter() - t0) * 1000, 2) snap = get_process_vram_bytes() _record({ "event_type": "ComfyUI VRAM Purge", "source": "ComfyUI Pipeline", "target": "VRAM Free", "duration_ms": duration_ms, "cache_status": "Cleaned", }) return {"success": True, "duration_ms": duration_ms, "free_vram_gb": round(snap["free_bytes"] / (1024**3), 2)} except Exception as e: return {"success": False, "error": str(e), "duration_ms": round((time.perf_counter() - t0) * 1000, 2)} _MODEL_SIZE_CACHE: Dict[str, int] = {} def _model_size_bytes(model_name: str) -> int: """On-disk weight size for an Ollama model, used to turn load time into bandwidth.""" if model_name in _MODEL_SIZE_CACHE: return _MODEL_SIZE_CACHE[model_name] try: for f in ram_optimizer.find_ollama_model_files(): _MODEL_SIZE_CACHE[f["model"]] = f["size_bytes"] except Exception as e: logger.debug(f"model size lookup failed: {e}") return _MODEL_SIZE_CACHE.get(model_name, 0) def classify_load(size_bytes: int, load_duration_ms: float) -> Dict[str, Any]: """Classify how a model reached VRAM, from achieved bandwidth rather than a constant. The old rule was `load_duration_ms < 2500`, which called a 27B Q2_K read from NVMe a cache hit and a small model read from RAM a cold load. Bandwidth separates them cleanly: page cache feeds PCIe at many GB/s, this NVMe does not. """ if load_duration_ms <= 1.0: return {"cache_status": "Already in VRAM", "load_gbps": None, "is_ram_hit": True} if not size_bytes: # No size on record — fall back to the old heuristic, but say so. # Without a size we cannot compute bandwidth at all; this is a guess and is # labelled as one. 8s roughly splits the measured warm (4.9s) and cold (34.3s) # loads for a mid-size model, but it is meaningless for very small or large ones. return { "cache_status": "RAM Cache Hit ⚡" if load_duration_ms < 8000 else "Cold Disk Load 💾", "load_gbps": None, "is_ram_hit": load_duration_ms < 8000, "detail": "size unknown, fell back to a duration guess", } gbps = (size_bytes / (1024**3)) / (load_duration_ms / 1000.0) if gbps >= RAM_HIT_GBPS: status = "RAM Cache Hit ⚡" elif gbps >= PARTIAL_HIT_GBPS: status = "Partial Cache 🌤" else: status = "Cold Disk Load 💾" return {"cache_status": status, "load_gbps": round(gbps, 2), "is_ram_hit": gbps >= RAM_HIT_GBPS} async def switch_ollama_model(target_model: str, keep_alive: str = "30m", _retrying: bool = False) -> Dict[str, Any]: """High-speed hot-swap to target Ollama model, tracking swap metrics. If the load fails because the model will not fit, reclaims VRAM from an idle ComfyUI and retries once. `_retrying` guards against recursing more than one level. """ t0 = time.perf_counter() cur_state = await get_ollama_live_state() prev_model = cur_state.get("active_model_name") or "None" try: client = _client(OLLAMA_API_BASE, 180.0) resp = await client.post( "/api/generate", json={"model": target_model, "prompt": "Ready check", "stream": False, "keep_alive": keep_alive}, ) total_duration_ms = round((time.perf_counter() - t0) * 1000, 2) if resp.status_code == 200: data = resp.json() load_dur_ms = round(data.get("load_duration", 0) / 1e6, 2) eval_dur_ms = round(data.get("eval_duration", 0) / 1e6, 2) eval_count = data.get("eval_count", 0) tokens_per_sec = round((eval_count / (eval_dur_ms / 1000)) if eval_dur_ms > 0 else 0, 1) size_bytes = _model_size_bytes(target_model) cls = classify_load(size_bytes, load_dur_ms) _record({ "event_type": "LLM Model Switch", "source": prev_model, "target": target_model, "duration_ms": total_duration_ms, "load_duration_ms": load_dur_ms, "tokens_per_sec": tokens_per_sec, "bytes_loaded": size_bytes, "load_gbps": cls["load_gbps"], "cache_status": cls["cache_status"], "detail": cls.get("detail"), }) return { "success": True, "prev_model": prev_model, "target_model": target_model, "total_duration_ms": total_duration_ms, "load_duration_ms": load_dur_ms, "tokens_per_sec": tokens_per_sec, "model_size_gb": round(size_bytes / (1024**3), 2) if size_bytes else None, "load_gbps": cls["load_gbps"], "cache_status": cls["cache_status"], "is_ram_hit": cls["is_ram_hit"], "response": data.get("response", ""), } # A model that will not fit is the exact contention this service exists to # resolve. Rather than handing the caller a CUDA OOM, take the VRAM back from an # idle ComfyUI and try once more. body = resp.text if looks_like_vram_oom(body) and not _retrying: snap = get_process_vram_bytes() if snap["comfyui_bytes"] >= RECLAIM_MIN_COMFY_BYTES: logger.warning( f"Ollama could not fit '{target_model}' with ComfyUI holding " f"{round(snap['comfyui_bytes'] / (1024**3), 2)} GB — reclaiming and retrying") purge = await instant_free_comfyui_vram() arbitrator.stats["reclaims_for_ollama"] += 1 arbitrator.last_action = ( f"Reclaimed {round(snap['comfyui_bytes'] / (1024**3), 2)}GB from ComfyUI so " f"'{target_model}' could load") _record({ "event_type": "VRAM Reclaim for Ollama", "source": "ComfyUI Pipeline", "target": target_model, "duration_ms": purge.get("duration_ms"), "cache_status": "Reclaimed", "detail": f"Ollama OOM: {body[:160]}", }) await asyncio.sleep(0.3) retry = await switch_ollama_model(target_model, keep_alive, _retrying=True) retry["reclaimed_from_comfyui_gb"] = round( snap["comfyui_bytes"] / (1024**3), 2) retry["first_attempt_error"] = "CUDA OOM; retried after reclaiming VRAM" return retry return {"success": False, "error": f"HTTP {resp.status_code}: {body}", "duration_ms": total_duration_ms, "vram_oom": looks_like_vram_oom(body)} except Exception as e: return {"success": False, "error": str(e), "duration_ms": round((time.perf_counter() - t0) * 1000, 2)} def get_switch_history() -> List[Dict[str, Any]]: return list(SWITCH_HISTORY) class AutoArbitrator: """Real-time bidirectional background arbitrator for seamless Ollama <-> ComfyUI hot-swapping. Two behavioural changes worth knowing about: * ComfyUI's VRAM is no longer purged 1.5 s after every finished prompt. Iterating on a workflow is the common case, and purging between runs forced a full checkpoint reload each time. The purge now waits for COMFY_IDLE_PURGE_S of genuinely empty queue, and happens immediately only when Ollama actually needs the VRAM. * The watchdog no longer polls two ComfyUI endpoints every 300 ms. The WebSocket is the primary signal; polling is a fallback that runs at 1 Hz and only hits /queue, backing off further while the socket is healthy. """ COMFY_IDLE_PURGE_S = 30.0 WATCHDOG_INTERVAL_S = 1.0 WATCHDOG_INTERVAL_WS_OK_S = 3.0 def __init__(self): self.running = False self.ws_task: Optional[asyncio.Task] = None self.poll_task: Optional[asyncio.Task] = None self.idle_task: Optional[asyncio.Task] = None self.last_yield_time = 0.0 self.last_comfy_free_time = 0.0 self.connected_ws = False self.last_action = "Idle" self.comfy_was_active = False self.comfy_idle_since: Optional[float] = None self.oc_profile = None self.pending_purge = False # While a tuning sweep is running, the arbitrator must not fight it: a ComfyUI # benchmark would otherwise trip trigger_comfy_priority, which reapplies the whole # 'comfy' profile and silently overwrites the clock the sweep is measuring. self.oc_suspended = False # Per-model backoff. A model that is mid-generation cannot yield, and asking it # again every second just blocks the loop repeatedly for no benefit. self._yield_backoff_until: Dict[str, float] = {} self._yield_busy_streak: Dict[str, int] = {} self.last_reclaim_time = 0.0 self.stats = { "yields": 0, # release confirmed "yield_deferred_busy": 0, # model mid-generation; unload queued behind it "yield_stalled": 0, # VRAM held with an idle GPU -- the real failure "deferred_releases": 0, # queued unloads that later landed "purges": 0, "deferred_purges": 0, "reclaims_for_ollama": 0, # ComfyUI purged because the LLM was spilling to CPU } async def start(self): if self.running: return self.running = True self.ws_task = asyncio.create_task(self._ws_listener()) self.poll_task = asyncio.create_task(self._poll_watchdog()) self.idle_task = asyncio.create_task(self._idle_purge_loop()) logger.info("AutoArbitrator background engine started (Bidirectional).") try: await asyncio.get_running_loop().run_in_executor( None, overclock_manager.apply_profile, "balanced") self.oc_profile = "balanced" except Exception as e: logger.warning(f"Startup overclock apply failed: {e}") async def stop(self): self.running = False for task in (self.ws_task, self.poll_task, self.idle_task): if task: task.cancel() await close_clients() logger.info("AutoArbitrator background engine stopped.") # Backoff schedule for a model that keeps reporting busy, in seconds. BUSY_BACKOFF_S = (5.0, 15.0, 30.0, 60.0) def note_deferred_release(self, waited_ms: float) -> None: """Called when a queued unload finally lands after a generation finished.""" self.stats["deferred_releases"] += 1 self._yield_backoff_until.clear() self._yield_busy_streak.clear() self.last_action = (f"VRAM released after the LLM finished " f"({round(waited_ms / 1000, 1)}s) — ComfyUI can proceed") async def trigger_comfy_priority(self, reason: str = "ComfyUI prompt detected"): """Yield Ollama's VRAM before diffusion allocates, without fighting a busy model.""" self.comfy_was_active = True self.comfy_idle_since = None self._apply_oc_profile("comfy") now = time.time() if now - self.last_yield_time < 1.0: return ollama_state = await get_ollama_live_state() model = ollama_state.get("active_model_name") if not model: return # Still finishing an inference we already asked to unload: leave it alone. until = self._yield_backoff_until.get(model, 0.0) if now < until: return logger.info(f"⚡ ComfyUI active ({reason}) -> Auto-yielding Ollama model '{model}'...") self.last_yield_time = time.time() res = await instant_free_ollama_vram(model, confirm=True) outcome = res.get("outcome") if outcome == "released": self.stats["yields"] += 1 self._yield_backoff_until.pop(model, None) self._yield_busy_streak.pop(model, None) self.last_action = (f"Yielded '{model}' for ComfyUI in " f"{res.get('confirm_ms')}ms ({res.get('freed_gb')}GB freed)") elif outcome == "busy": streak = self._yield_busy_streak.get(model, 0) delay = self.BUSY_BACKOFF_S[min(streak, len(self.BUSY_BACKOFF_S) - 1)] self._yield_busy_streak[model] = streak + 1 self._yield_backoff_until[model] = time.time() + delay self.stats["yield_deferred_busy"] += 1 self.last_action = (f"'{model}' is mid-generation ({res.get('gpu_util_pct')}% GPU); " f"unload is queued and will apply when it finishes") logger.info(f"Yield deferred: {res.get('error')} — backing off {delay}s") else: self.stats["yield_stalled"] += 1 self.last_action = (f"⚠ '{model}' holding {res.get('residual_gb')}GB with an idle GPU") logger.warning(f"VRAM yield stalled: {res.get('error')}") async def trigger_comfy_completed(self, immediate: bool = False): """Mark the end of a generation. The actual purge is deferred unless forced.""" self.comfy_was_active = False if self.comfy_idle_since is None: self.comfy_idle_since = time.time() self._apply_oc_profile("ollama") if immediate: await self._purge_comfy_now("Ollama needs VRAM") else: self.pending_purge = True self.stats["deferred_purges"] += 1 self.last_action = (f"ComfyUI idle — holding its checkpoints for " f"{int(self.COMFY_IDLE_PURGE_S)}s in case you iterate") async def _purge_comfy_now(self, reason: str): now = time.time() if now - self.last_comfy_free_time < 3.0: return self.last_comfy_free_time = now self.pending_purge = False logger.info(f"⚡ Purging ComfyUI VRAM cache ({reason})...") res = await instant_free_comfyui_vram() self.stats["purges"] += 1 self.last_action = f"Purged ComfyUI VRAM ({res.get('duration_ms')}ms) — {reason}" logger.info(f"ComfyUI purge completed: {res}") async def _idle_purge_loop(self): """Purge ComfyUI's VRAM only after a real idle gap, not between iterations.""" while self.running: try: if self.pending_purge and self.comfy_idle_since and not self.comfy_was_active: idle_for = time.time() - self.comfy_idle_since if idle_for >= self.COMFY_IDLE_PURGE_S: await self._purge_comfy_now( f"idle {int(idle_for)}s") except Exception as e: logger.debug(f"idle purge loop error: {e}") await asyncio.sleep(2.0) async def request_vram_for_ollama(self, needed_gb: float = 0.0) -> Dict[str, Any]: """Called when Ollama needs VRAM now: purge ComfyUI immediately rather than waiting.""" snap = get_process_vram_bytes() free_gb = snap["free_bytes"] / (1024**3) if needed_gb and free_gb >= needed_gb: return {"purged": False, "free_gb": round(free_gb, 2), "reason": "enough free VRAM"} if snap["comfyui_bytes"] > YIELD_RESIDUAL_BYTES: await self._purge_comfy_now(f"Ollama requested {needed_gb or '?'}GB") snap = get_process_vram_bytes() return {"purged": True, "free_gb": round(snap["free_bytes"] / (1024**3), 2)} return {"purged": False, "free_gb": round(free_gb, 2), "reason": "ComfyUI holds no VRAM"} async def _ws_listener(self): client_id = "hyperswap-arbitrator" ws_url = f"ws://127.0.0.1:8188/ws?clientId={client_id}" backoff = 2.0 while self.running: try: async with websockets.connect(ws_url, ping_interval=10, ping_timeout=10) as ws: self.connected_ws = True backoff = 2.0 logger.info("AutoArbitrator connected to ComfyUI WebSocket.") while self.running: msg = await ws.recv() if not isinstance(msg, str): continue try: data = json.loads(msg) msg_type = data.get("type") msg_data = data.get("data", {}) if msg_type == "status": queue_rem = (msg_data.get("status", {}) .get("exec_info", {}).get("queue_remaining", 0)) if queue_rem > 0: await self.trigger_comfy_priority(f"Queue remaining: {queue_rem}") elif queue_rem == 0 and self.comfy_was_active: await self.trigger_comfy_completed() elif msg_type in ("execution_start", "execution_cached"): await self.trigger_comfy_priority(f"Event: {msg_type}") elif msg_type == "executing": node = msg_data.get("node") if node is not None: await self.trigger_comfy_priority(f"Executing node: {node}") elif self.comfy_was_active: await self.trigger_comfy_completed() elif msg_type == "execution_success": await self.trigger_comfy_completed() elif msg_type == "execution_error": logger.warning(f"ComfyUI execution error: {msg_data}") await self.trigger_comfy_completed() except Exception as e: logger.debug(f"WS parse error: {e}") except (websockets.exceptions.ConnectionClosed, OSError, asyncio.CancelledError): self.connected_ws = False except Exception as e: self.connected_ws = False logger.debug(f"WS connection error: {e}") await asyncio.sleep(backoff) backoff = min(backoff * 1.5, 15.0) RECLAIM_COOLDOWN_S = 30.0 async def _check_ollama_starved(self) -> None: """The other direction: rescue an LLM that ComfyUI has squeezed onto the CPU. Yielding Ollama for ComfyUI was automatic; the reverse never was, despite the README calling the arbitration bidirectional. When Ollama cannot fit a model it does not fail, it silently places layers on the CPU and runs about an order of magnitude slower -- so this is the failure mode a user is least likely to notice and most likely to feel. If the LLM is spilling while ComfyUI sits idle holding VRAM, ComfyUI's cached checkpoints are the thing to give up. """ now = time.time() if self.comfy_was_active or (now - self.last_reclaim_time) < self.RECLAIM_COOLDOWN_S: return ollama = await get_ollama_live_state() if not ollama.get("partially_offloaded"): return snap = get_process_vram_bytes() if snap["comfyui_bytes"] < RECLAIM_MIN_COMFY_BYTES: return # ComfyUI is not the one holding the memory; nothing we can do here self.last_reclaim_time = now model = ollama.get("active_model_name") offload = ollama.get("cpu_offload_pct") logger.warning(f"⚠ '{model}' is {offload}% on CPU while ComfyUI holds " f"{round(snap['comfyui_bytes'] / (1024**3), 2)} GB — reclaiming for the LLM") await self._purge_comfy_now(f"LLM spilling {offload}% to CPU") self.stats["reclaims_for_ollama"] += 1 # Freeing VRAM does not move layers back; only a reload re-places the model. Do # that only when the model is idle, never mid-generation. after = get_process_vram_bytes() if after.get("gpu_util_pct", 0) < BUSY_UTIL_PCT and model: logger.info(f"Reloading '{model}' to place it fully on the GPU...") await instant_free_ollama_vram(model, confirm=True) res = await switch_ollama_model(model, keep_alive="30m") recheck = await get_ollama_live_state() self.last_action = ( f"Reclaimed {round(snap['comfyui_bytes'] / (1024**3), 2)}GB from ComfyUI and " f"reloaded '{model}' — now {round(recheck.get('gpu_fraction', 0) * 100)}% on GPU" if res.get("success") else f"Reclaimed VRAM from ComfyUI but reloading '{model}' failed: {res.get('error')}") else: self.last_action = (f"Reclaimed VRAM from ComfyUI; '{model}' is busy, so it will " f"stay partly on CPU until its next load") async def _poll_watchdog(self): """Fallback for when the WebSocket is down. One cheap /queue call, 1 Hz. The previous version hit /system_stats and /queue every 300 ms on fresh TCP connections — roughly 6.6 requests/second against ComfyUI, forever. """ while self.running: interval = self.WATCHDOG_INTERVAL_WS_OK_S if self.connected_ws else self.WATCHDOG_INTERVAL_S try: client = _client(COMFY_API_BASE, 3.0) resp = await client.get("/queue") if resp.status_code == 200: q = resp.json() busy = len(q.get("queue_running", [])) > 0 or len(q.get("queue_pending", [])) > 0 if busy: await self.trigger_comfy_priority("Watchdog saw an active queue") elif self.comfy_was_active: await self.trigger_comfy_completed() else: await self._check_ollama_starved() except Exception: pass await asyncio.sleep(interval) def suspend_oc(self, reason: str = "tuning sweep") -> None: self.oc_suspended = True logger.info(f"Overclock auto-switching suspended ({reason})") def resume_oc(self, profile: Optional[str] = None) -> None: self.oc_suspended = False # Forget the cached profile so the next transition actually reapplies. self.oc_profile = profile logger.info("Overclock auto-switching resumed") def _apply_oc_profile(self, profile: str): """Apply an overclock profile in a background thread; only fire on transition.""" if self.oc_suspended or self.oc_profile == profile: return self.oc_profile = profile try: loop = asyncio.get_running_loop() loop.run_in_executor(None, overclock_manager.apply_profile, profile) logger.info(f"🎛️ Overclock profile switched -> '{profile}'") except Exception as e: logger.warning(f"Overclock profile switch failed ({profile}): {e}") def get_status(self) -> Dict[str, Any]: idle_for = (time.time() - self.comfy_idle_since) if self.comfy_idle_since else None return { "running": self.running, "connected_ws": self.connected_ws, "last_action": self.last_action, "mode": "Bidirectional Hot-Swap (ComfyUI <-> Ollama)", "comfy_active": self.comfy_was_active, "pending_purge": self.pending_purge, "comfy_idle_s": round(idle_for, 1) if idle_for is not None else None, "idle_purge_after_s": self.COMFY_IDLE_PURGE_S, "oc_profile": self.oc_profile, "counters": dict(self.stats), "yield_backoff": {m: round(max(t - time.time(), 0), 1) for m, t in self._yield_backoff_until.items() if t > time.time()}, } arbitrator = AutoArbitrator()