ROCm-Research-Archive/_TestScripts/Scripts and Tests/bc250_softmax_patch_v13.py

"""
BC-250 gfx1010 Comprehensive Monkey-Patch v17

1. BF16 KILL: gfx1010 has NO native bf16 — force f16 everywhere
2. Softmax: manual impl for dim > threshold (VGPR overflow fix)
3. SDPA: manual impl for large sequences
4. GGUF: GGMLTensor.to() patched — quantized weights ALWAYS stay on CPU
   CPU dequant → f16 → GPU transfer per layer (gfx1010 GPU can't dequant)
5. Mmap: pre-clones non-GGUF tensor data before GPU transfer
6. GPU: pre-warms context and caching allocator
7. CLIP: forces text encoder to CPU (memory constraint)
8. VAE: CPU f32 decode cached in RAM (faster than GPU on this APU)
9. Threads: all CPU cores for intra-op parallelism
10. Rope: force rope() to CPU — gfx1010 has NO native float64
11. Non-blocking disabled: gfx1010 without SDMA hangs on async copies

v14: rope() → CPU (gfx1010 has no float64 HW)
v15: GGUF dequant → CPU (gfx1010 GPU hangs on Q5_K bitwise ops)
v16: GGMLTensor.to() patched to keep quantized weights on CPU
v17: cast_to → direct .to(non_blocking=False), warmup removed
  - non_blocking=True hangs on gfx1010 (no SDMA, async HIP copy broken)
  - replaced empty_like+copy_ with direct .to() for CPU→GPU transfer
  - device_supports_non_blocking → always False for this device

BC-250 APU / gfx1010: XNACK disabled, no GPU page fault handling.
Place in ComfyUI root and import as first line of main.py.
"""
import os
os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")

import torch
import torch.nn.functional as F
import sys
import gc
import logging
import time as _time

logger = logging.getLogger(__name__)

# === THREAD CONFIGURATION ===
_NUM_THREADS = int(os.environ.get("BC250_NUM_THREADS", str(os.cpu_count() or 12)))
torch.set_num_threads(_NUM_THREADS)
logger.warning(f"[BC-250] Torch threads: intra-op={_NUM_THREADS}")

# Disable torch._dynamo — gfx1010 doesn't benefit, compilation overhead is massive
try:
    torch._dynamo.config.suppress_errors = True
    logger.warning("[BC-250] torch._dynamo: TORCHDYNAMO_DISABLE=1 + suppress_errors")
except Exception:
    logger.warning("[BC-250] torch._dynamo: TORCHDYNAMO_DISABLE=1 (env only)")


SAFE_SOFTMAX_THRESHOLD = int(os.environ.get("BC250_SOFTMAX_THRESHOLD", "4096"))

_original_softmax = torch.nn.functional.softmax
_original_tensor_softmax = torch.Tensor.softmax
_original_sdpa = torch.nn.functional.scaled_dot_product_attention

# === MMAP PRE-CLONE + BF16 KILL PATCH ===
_original_module_apply = torch.nn.Module._apply

def _is_ggml_tensor(t):
    """Check if tensor is a GGMLTensor (has GGUF quantization metadata)."""
    return hasattr(t, 'tensor_type')

# === GGML TENSOR CPU LOCK ===
# Patched later when GGUF module loads (_try_patch_gguf).
# GGMLTensor.to() is monkey-patched so quantized weights NEVER leave CPU.
# This prevents both: GPU dequant hang AND GPU→CPU transfer hang.

def _bc250_safe_apply(self, fn, recurse=True):
    """Pre-clone mmap'd CPU tensor data before GPU transfer (XNACK workaround).
    Converts ALL bf16 → f16 (gfx1010 has no native bf16 — including GGML BF16).
    BF16 GGML tensors are dequantized to f32→f16, becoming regular tensors.
    Note: GGMLTensor.clone() returns self, so quantized GGUF weights are unaffected."""
    for key, param in self._parameters.items():
        if param is None:
            continue
        # Clone CPU data for XNACK workaround (skip GGML: clone() returns self)
        if param.device.type == 'cpu' and not _is_ggml_tensor(param.data):
            param.data = param.data.clone()
        # gfx1010: no native bf16. Convert ALL bf16 → f16 (including GGML BF16)
        if param.data.dtype == torch.bfloat16:
            param.data = param.data.float().half()
    for key, buf in self._buffers.items():
        if buf is None:
            continue
        if buf.device.type == 'cpu' and not _is_ggml_tensor(buf):
            buf = buf.clone()
        if buf.dtype == torch.bfloat16:
            buf = buf.float().half()
        self._buffers[key] = buf
    return _original_module_apply(self, fn, recurse)

# === SOFTMAX PATCH ===

def _safe_softmax_impl(input, dim=-1):
    x_max = input.max(dim=dim, keepdim=True).values
    exp_x = torch.exp(input - x_max)
    return exp_x / exp_x.sum(dim=dim, keepdim=True)

def patched_softmax(input, dim=None, _stacklevel=3, dtype=None):
    if dim is None:
        dim = -1
    if dtype is not None:
        input = input.to(dtype)
    if input.is_cuda and input.shape[dim] > SAFE_SOFTMAX_THRESHOLD:
        if not getattr(patched_softmax, '_logged', False):
            logger.warning(f"[BC-250] Manual F.softmax: shape={list(input.shape)}, dim={dim}")
            patched_softmax._logged = True
        return _safe_softmax_impl(input, dim)
    return _original_softmax(input, dim=dim)

def patched_tensor_softmax(self, dim=-1, dtype=None):
    if dtype is not None:
        self = self.to(dtype)
    if self.is_cuda and self.shape[dim] > SAFE_SOFTMAX_THRESHOLD:
        if not getattr(patched_tensor_softmax, '_logged', False):
            logger.warning(f"[BC-250] Manual softmax: shape={list(self.shape)}, dim={dim}")
            patched_tensor_softmax._logged = True
        return _safe_softmax_impl(self, dim)
    return _original_tensor_softmax(self, dim=dim)

def _safe_sdpa(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None):
    L, S = query.size(-2), key.size(-2)
    if scale is None:
        scale = query.size(-1) ** -0.5
    attn_weight = torch.matmul(query, key.transpose(-2, -1)) * scale
    if is_causal:
        causal_mask = torch.triu(torch.ones(L, S, dtype=torch.bool, device=query.device), diagonal=1)
        attn_weight = attn_weight.masked_fill(causal_mask, float('-inf'))
    if attn_mask is not None:
        if attn_mask.dtype == torch.bool:
            attn_weight = attn_weight.masked_fill(~attn_mask, float('-inf'))
        else:
            attn_weight = attn_weight + attn_mask
    attn_weight = _safe_softmax_impl(attn_weight, dim=-1)
    if dropout_p > 0.0:
        attn_weight = torch.nn.functional.dropout(attn_weight, p=dropout_p)
    return torch.matmul(attn_weight, value)

def patched_sdpa(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None):
    S = key.size(-2)
    if query.is_cuda and S > SAFE_SOFTMAX_THRESHOLD:
        if not getattr(patched_sdpa, '_logged', False):
            logger.warning(f"[BC-250] Manual SDPA: Q={list(query.shape)}, S={S}")
            patched_sdpa._logged = True
        return _safe_sdpa(query, key, value, attn_mask=attn_mask,
                         dropout_p=dropout_p, is_causal=is_causal, scale=scale)
    return _original_sdpa(query, key, value, attn_mask=attn_mask,
                         dropout_p=dropout_p, is_causal=is_causal, scale=scale)

# === GGUF CPU DEQUANT + GGMLTensor CPU LOCK ===
_gguf_patched = False

def _try_patch_gguf():
    """Patch GGUF for BC-250:
    1. GGMLTensor.to() → keeps quantized weights on CPU (ignore device arg)
    2. cast_bias_weight → CPU dequant, then transfer f16 result to GPU
    
    gfx1010 GPU cannot dequantize Q5_K (bitwise ops hang).
    And once weights are on GPU, transferring back to CPU also hangs.
    Only safe path: weights stay CPU → dequant on CPU → f16 to GPU."""
    global _gguf_patched
    if _gguf_patched:
        return True

    ops_mod = None
    dequant_mod = None
    for name, mod in sys.modules.items():
        if mod is None:
            continue
        if name.endswith('.ops') and 'GGUF' in name:
            ops_mod = mod
        if name.endswith('.dequant') and 'GGUF' in name:
            dequant_mod = mod

    if ops_mod is None or dequant_mod is None:
        return False

    GGMLLayer = getattr(ops_mod, 'GGMLLayer', None)
    GGMLTensor = getattr(ops_mod, 'GGMLTensor', None)
    is_quantized_fn = getattr(dequant_mod, 'is_quantized', None)
    if GGMLLayer is None or GGMLTensor is None or is_quantized_fn is None:
        return False

    torch_compiler_disable = getattr(ops_mod, 'torch_compiler_disable', None)

    # === PATCH 1: GGMLTensor.to() — keep quantized on CPU ===
    _original_ggml_to = GGMLTensor.to

    def _bc250_ggml_to(self, *args, **kwargs):
        """Intercept .to() calls: keep quantized weights on CPU.
        Only allow dtype changes, block device changes to CUDA.
        This prevents load_models_gpu from moving GGUF weights to GPU."""
        # Check if this is a quantized tensor (has tensor_type metadata)
        if hasattr(self, 'tensor_type') and self.tensor_type is not None:
            # Parse the .to() call to extract device, dtype, non_blocking
            # Common patterns from nn.Module._apply:
            #   t.to(device, dtype, non_blocking)  — 3 positional
            #   t.to(device)                       — 1 positional
            #   t.to(dtype)                        — 1 positional (dtype)
            #   t.to(device=..., dtype=..., non_blocking=...)  — kwargs
            parsed_device = kwargs.get('device', None)
            parsed_dtype = kwargs.get('dtype', None)
            parsed_nb = kwargs.get('non_blocking', False)
            parsed_mem_fmt = kwargs.get('memory_format', None)

            for a in args:
                if isinstance(a, torch.device):
                    parsed_device = a
                elif isinstance(a, str):
                    try:
                        parsed_device = torch.device(a)
                    except Exception:
                        pass
                elif isinstance(a, torch.dtype):
                    parsed_dtype = a
                elif isinstance(a, bool):
                    parsed_nb = a
                elif a is None:
                    # dtype=None from Module.to() convert function
                    pass

            # Block CUDA transfer for quantized weights — stay on CPU
            if parsed_device is not None and parsed_device.type == 'cuda':
                # Reconstruct call without device, keeping dtype/non_blocking
                remap_kwargs = {}
                if parsed_dtype is not None:
                    remap_kwargs['dtype'] = parsed_dtype
                if parsed_nb:
                    remap_kwargs['non_blocking'] = parsed_nb
                if parsed_mem_fmt is not None:
                    remap_kwargs['memory_format'] = parsed_mem_fmt
                if remap_kwargs:
                    return _original_ggml_to(self, **remap_kwargs)
                return self  # No-op: was just a device move

        return _original_ggml_to(self, *args, **kwargs)

    GGMLTensor.to = _bc250_ggml_to
    logger.warning("[BC-250] GGMLTensor.to() patched: quantized weights locked to CPU")

    # === PATCH 2: cast_bias_weight — CPU dequant + GPU transfer ===
    _fwd_count = [0, 0.0]

    def _bc250_cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
        """CPU dequant → GPU transfer. Weights are guaranteed CPU (GGMLTensor.to patched).
        Dequant on CPU via get_weight(), transfer f16 to GPU with synchronous .to().
        gfx1010 without SDMA cannot do async copies — non_blocking=False always."""
        if _fwd_count[0] == 0:
            _fwd_count[1] = _time.time()
        _fwd_count[0] += 1

        if input is not None:
            if dtype is None:
                dtype = getattr(input, "dtype", torch.float32)
            if bias_dtype is None:
                bias_dtype = dtype
            if device is None:
                device = input.device

        # gfx1010: never dequant to bf16
        if dtype == torch.bfloat16:
            dtype = torch.float16
        if bias_dtype == torch.bfloat16:
            bias_dtype = torch.float16

        is_first = _fwd_count[0] <= 3 or 499 <= _fwd_count[0] <= 505
        if is_first:
            logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: device={device}, w_type={type(s.weight).__name__}, w_dev={s.weight.device}, has_tt={hasattr(s.weight, 'tensor_type')}")

        bias = None
        if s.bias is not None:
            if is_first:
                logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: get_weight(bias)...")
            bias = s.get_weight(s.bias, bias_dtype)
            if is_first:
                logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: bias got, type={type(bias).__name__}, dev={bias.device}, dt={bias.dtype}")
            if type(bias) is not torch.Tensor:
                bias = bias.as_subclass(torch.Tensor)
            if bias.device != device or bias.dtype != bias_dtype:
                if is_first:
                    logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: bias.to({device})...")
                bias = bias.to(device=device, dtype=bias_dtype, non_blocking=False)
                if is_first:
                    logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: bias transferred")

        if is_first:
            logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: get_weight(weight)...")
        weight = s.get_weight(s.weight, dtype)
        if is_first:
            logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: weight got, type={type(weight).__name__}, shape={list(weight.shape)}, dev={weight.device}, dt={weight.dtype}")
        if type(weight) is not torch.Tensor:
            weight = weight.as_subclass(torch.Tensor)
        if weight.device != device or weight.dtype != dtype:
            if is_first:
                logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: weight.to({device}, {dtype})...")
            weight = weight.to(device=device, dtype=dtype, non_blocking=False)
            if is_first:
                logger.warning(f"[BC-250] CBW#{_fwd_count[0]}: weight transferred")

        if _fwd_count[0] % 100 == 0:
            elapsed = _time.time() - _fwd_count[1]
            logger.warning(f"[BC-250] Layer {_fwd_count[0]}, elapsed {elapsed:.1f}s")

        # Log all CUDA transfers to find the exact hang point
        if device is not None and hasattr(device, 'type') and device.type == 'cuda':
            logger.warning(f"[BC-250] CUDA#{_fwd_count[0]}: {list(weight.shape)} {weight.dtype} done")

        return weight, bias

    if torch_compiler_disable is not None:
        _bc250_cast_bias_weight = torch_compiler_disable()(_bc250_cast_bias_weight)

    GGMLLayer.cast_bias_weight = _bc250_cast_bias_weight
    _gguf_patched = True
    logger.warning("[BC-250] GGUF patched: CPU dequant + GGMLTensor CPU-locked")
    return True

class _GGUFImportWatcher:
    def __init__(self):
        self.done = False
    def find_module(self, fullname, path=None):
        if self.done:
            return None
        if 'GGUF' in fullname and ('dequant' in fullname or 'ops' in fullname):
            return self
        return None
    def load_module(self, fullname):
        if self in sys.meta_path:
            sys.meta_path.remove(self)
        try:
            import importlib
            mod = importlib.import_module(fullname)
        finally:
            if self not in sys.meta_path:
                sys.meta_path.insert(0, self)
        if _try_patch_gguf():
            self.done = True
        return mod

# === MODEL MANAGEMENT PATCHES (text encoder CPU + bf16 kill + load hook) ===
_mm_patched = False

def _try_patch_model_management():
    """Patches comfy.model_management:
    - Text encoder → CPU
    - should_use_bf16 → always False
    - unet_dtype → never returns bf16
    - load_models_gpu → cleanup + post-load bf16→f16
    """
    global _mm_patched
    if _mm_patched:
        return True

    mm = sys.modules.get('comfy.model_management')
    if mm is None:
        return False

    # Text encoder on CPU
    mm.text_encoder_device = lambda: torch.device("cpu")
    mm.text_encoder_offload_device = lambda: torch.device("cpu")
    logger.warning("[BC-250] Text encoder forced to CPU")

    # Kill bf16 globally — gfx1010 has no native bf16
    mm.should_use_bf16 = lambda *a, **kw: False
    logger.warning("[BC-250] should_use_bf16 → always False (gfx1010)")

    # Force non_blocking=False — gfx1010 without SDMA hangs on async HIP copies
    mm.device_supports_non_blocking = lambda *a, **kw: False
    logger.warning("[BC-250] device_supports_non_blocking → always False (no SDMA)")

    _original_unet_dtype = mm.unet_dtype
    def _bc250_unet_dtype(*args, **kwargs):
        return torch.float16  # gfx1010: always f16 (2× faster than f32, no bf16 HW)
    mm.unet_dtype = _bc250_unet_dtype
    logger.warning("[BC-250] unet_dtype patched: always f16")

    # Ensure fp16 is recognized as available
    mm.should_use_fp16 = lambda *a, **kw: True
    logger.warning("[BC-250] should_use_fp16 → always True")

    # Load hook: cleanup + post-load bf16 → f16 conversion
    _original_load = getattr(mm, 'load_models_gpu', None)
    if _original_load is not None:
        def _bc250_load_models_gpu(models, *args, **kwargs):
            gc.collect()
            torch.cuda.empty_cache()
            result = _original_load(models, *args, **kwargs)
            # Post-load: convert ALL remaining bf16 params/buffers to f16
            for m in models:
                real_model = getattr(m, 'model', None)
                if real_model is None:
                    continue
                converted = 0
                for p in real_model.parameters():
                    if p.dtype == torch.bfloat16:
                        p.data = p.data.float().half()
                        converted += 1
                for name, buf in real_model.named_buffers():
                    if buf is not None and buf.dtype == torch.bfloat16:
                        parts = name.split('.')
                        obj = real_model
                        for part in parts[:-1]:
                            obj = getattr(obj, part)
                        setattr(obj, parts[-1], buf.float().half())
                        converted += 1
                if converted > 0:
                    logger.warning(f"[BC-250] Post-load: converted {converted} bf16→f16 params/buffers")
            return result
        mm.load_models_gpu = _bc250_load_models_gpu
        logger.warning("[BC-250] GPU load hook installed (cleanup + bf16 kill)")

    _mm_patched = True
    return True

class _ModelMgmtWatcher:
    def __init__(self):
        self.done = False
    def find_module(self, fullname, path=None):
        if self.done:
            return None
        if fullname == 'comfy.model_management':
            return self
        return None
    def load_module(self, fullname):
        if self in sys.meta_path:
            sys.meta_path.remove(self)
        try:
            import importlib
            mod = importlib.import_module(fullname)
        finally:
            if self not in sys.meta_path:
                sys.meta_path.insert(0, self)
        if _try_patch_model_management():
            self.done = True
        return mod

# === VAE CPU CACHED ===
_vae_patched = False
_vae_cached = False

def _try_patch_vae_gpu():
    global _vae_patched
    if _vae_patched:
        return True

    sd_mod = sys.modules.get('comfy.sd')
    if sd_mod is None:
        return False

    VAE = getattr(sd_mod, 'VAE', None)
    if VAE is None:
        return False

    def _ensure_vae_cached_cpu(self):
        """Keep VAE on CPU in RAM, eval mode. No GPU transfer needed.
        On BC-250 APU: GPU VAE decode is slower than CPU (24 CUs, no SDMA).
        CPU has 12 Zen2 threads and direct RAM access — faster for VAE convolutions."""
        global _vae_cached
        if not _vae_cached:
            t0 = _time.time()
            self.first_stage_model.to(device='cpu', dtype=torch.float32)
            self.first_stage_model.eval()
            _vae_cached = True
            logger.warning(f"[BC-250] VAE cached on CPU (f32) in {_time.time()-t0:.1f}s")
        self.disable_offload = True

    def _bc250_vae_decode(self, samples_in, vae_options={}):
        t0 = _time.time()
        self.throw_exception_if_invalid()

        if self.latent_dim == 2 and samples_in.ndim == 5:
            samples_in = samples_in[:, :, 0]

        _ensure_vae_cached_cpu(self)

        pixel_samples = None
        with torch.no_grad():
            for x in range(samples_in.shape[0]):
                sample = samples_in[x:x+1].float().cpu()
                decoded = self.first_stage_model.decode(sample, **vae_options)
                if decoded.ndim == 5:
                    decoded = decoded[:, :, 0]
                out = self.process_output(decoded.float())
                if pixel_samples is None:
                    pixel_samples = torch.empty(
                        (samples_in.shape[0],) + tuple(out.shape[1:]), device='cpu'
                    )
                pixel_samples[x:x+1] = out
                del decoded, sample

        pixel_samples = pixel_samples.movedim(1, -1)
        elapsed = _time.time() - t0
        logger.warning(f"[BC-250] VAE decode (CPU f32): {elapsed:.1f}s")
        return pixel_samples

    def _bc250_vae_encode(self, pixel_samples):
        t0 = _time.time()
        self.throw_exception_if_invalid()
        _ensure_vae_cached_cpu(self)
        with torch.no_grad():
            pixels_in = self.process_input(pixel_samples).float().cpu()
            result = self.first_stage_model.encode(pixels_in).float()
        logger.warning(f"[BC-250] VAE encode (CPU f32): {_time.time() - t0:.1f}s")
        return result

    VAE.decode = _bc250_vae_decode
    VAE.encode = _bc250_vae_encode
    _vae_patched = True
    logger.warning("[BC-250] VAE patched: CPU f32 decode/encode (cached in RAM)")
    return True

class _SDModuleWatcher:
    def __init__(self):
        self.done = False
    def find_module(self, fullname, path=None):
        if self.done:
            return None
        if fullname == 'comfy.sd':
            return self
        return None
    def load_module(self, fullname):
        if self in sys.meta_path:
            sys.meta_path.remove(self)
        try:
            import importlib
            mod = importlib.import_module(fullname)
        finally:
            if self not in sys.meta_path:
                sys.meta_path.insert(0, self)
        if _try_patch_vae_gpu():
            self.done = True
        return mod

# === ROPE CPU PATCH (gfx1010 has no float64 hardware) ===
_rope_patched = False

def _try_patch_rope():
    """Patch rope() in flux/math.py to always compute on CPU.
    gfx1010 has no native float64 — GPU float64 ops are software-emulated and hang."""
    global _rope_patched
    if _rope_patched:
        return True

    flux_math = sys.modules.get('comfy.ldm.flux.math')
    if flux_math is None:
        return False

    _original_rope = getattr(flux_math, 'rope', None)
    if _original_rope is None:
        return False

    def _bc250_rope(pos, dim, theta):
        """Compute rope on CPU (float64 not supported on gfx1010), then move result to original device."""
        assert dim % 2 == 0
        target_device = pos.device
        device = torch.device("cpu")
        scale = torch.linspace(0, (dim - 2) / dim, steps=dim // 2, dtype=torch.float64, device=device)
        omega = 1.0 / (theta ** scale)
        out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega)
        from einops import rearrange
        out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
        out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
        return out.to(dtype=torch.float32, device=target_device)

    flux_math.rope = _bc250_rope
    _rope_patched = True
    logger.warning("[BC-250] rope() patched: CPU computation (no float64 on gfx1010)")
    return True

class _FluxMathWatcher:
    def __init__(self):
        self.done = False
    def find_module(self, fullname, path=None):
        if self.done:
            return None
        if fullname == 'comfy.ldm.flux.math':
            return self
        return None
    def load_module(self, fullname):
        if self in sys.meta_path:
            sys.meta_path.remove(self)
        try:
            import importlib
            mod = importlib.import_module(fullname)
        finally:
            if self not in sys.meta_path:
                sys.meta_path.insert(0, self)
        if _try_patch_rope():
            self.done = True
        return mod

# === WARMUP REMOVED (v17) ===
# First prompt may be slower, subsequent prompts benefit from HIP kernel caches.

# === INSTALL ===

def _prewarm_gpu():
    try:
        if not torch.cuda.is_available():
            return
        # Warm GPU context + allocator
        dummy = torch.zeros(1, device='cuda')
        _ = dummy + 1
        torch.cuda.synchronize()
        # Warm CPU→GPU copy kernel (COMGR JIT on first transfer)
        cpu_t = torch.randn(256, 256, dtype=torch.float16)
        gpu_t = cpu_t.to('cuda')
        _ = torch.matmul(gpu_t, gpu_t.T)
        torch.cuda.synchronize()
        del dummy, cpu_t, gpu_t, _
        torch.cuda.empty_cache()
        logger.warning("[BC-250] GPU pre-warmed (context + copy + matmul)")
    except Exception as e:
        logger.warning(f"[BC-250] GPU pre-warm failed: {e}")


def install():
    # Mmap pre-clone + bf16 kill patch
    torch.nn.Module._apply = _bc250_safe_apply
    logger.warning("[BC-250] Mmap pre-clone + bf16→f16 patch installed")

    # Softmax patches
    torch.nn.functional.softmax = patched_softmax
    torch.Tensor.softmax = patched_tensor_softmax
    torch.nn.functional.scaled_dot_product_attention = patched_sdpa
    logger.warning(f"[BC-250] Softmax patch installed (threshold={SAFE_SOFTMAX_THRESHOLD})")

    # Deferred patches via import hooks
    sys.meta_path.insert(0, _GGUFImportWatcher())
    sys.meta_path.insert(0, _ModelMgmtWatcher())
    sys.meta_path.insert(0, _SDModuleWatcher())
    sys.meta_path.insert(0, _FluxMathWatcher())

    # Try immediate patches
    _try_patch_gguf()
    _try_patch_model_management()
    _try_patch_vae_gpu()
    _try_patch_rope()

    _prewarm_gpu()
    logger.warning("[BC-250] v17 ready — no warmup, first prompt may be slow")

install()