460 lines
20 KiB
Python
460 lines
20 KiB
Python
import copy
|
|
import torch
|
|
import comfy.model_patcher
|
|
from comfy_api.latest import io
|
|
import comfy.hooks
|
|
import logging
|
|
import comfy.patcher_extension
|
|
|
|
from .Parser.UnifiedMathVisitor import UnifiedMathVisitor
|
|
from .helper_functions import checkLazyNew, generate_dim_variables, parse_expr, as_tensor
|
|
from .Stack import MrmthStack
|
|
from .ParseTree import MrmthParseTree
|
|
from comfy.ldm.modules.attention import optimized_attention
|
|
|
|
|
|
class SelectiveGuiderMathNode(io.ComfyNode):
|
|
@classmethod
|
|
def define_schema(cls) -> io.Schema:
|
|
return io.Schema(
|
|
node_id="mrmth_ag_SelectiveGuiderMathNode",
|
|
category="More math",
|
|
display_name="Selective guider math",
|
|
inputs=[
|
|
io.Guider.Input("V"),
|
|
io.Autogrow.Input(
|
|
id="F",
|
|
template=io.Autogrow.TemplatePrefix(
|
|
io.Float.Input("float", default=0.0, optional=True, lazy=True, force_input=True),
|
|
prefix="F",
|
|
min=1,
|
|
max=50,
|
|
),
|
|
),
|
|
io.MultiType.Input(
|
|
io.String.Input("Expression", default="push(0, inp); inp", multiline=False),
|
|
types=[io.String, MrmthParseTree],
|
|
),
|
|
io.Int.Input("layer_x", default=0, min=-999, max=999),
|
|
],
|
|
outputs=[io.Guider.Output(), MrmthStack.Output()],
|
|
)
|
|
|
|
@classmethod
|
|
def check_lazy_status(
|
|
cls,
|
|
V,
|
|
F,
|
|
Expression,
|
|
layer_x=0,
|
|
remember_stack=False,
|
|
stack={},
|
|
):
|
|
return checkLazyNew(Expression, {}, F)
|
|
|
|
@classmethod
|
|
def execute(
|
|
cls,
|
|
V,
|
|
F,
|
|
Expression,
|
|
layer_x=0,
|
|
remember_stack=False,
|
|
stack=None,
|
|
):
|
|
if V is None or not hasattr(V, "model_options"):
|
|
raise ValueError("Vstupní guider musí mít model_options.")
|
|
|
|
hook_target = "all"
|
|
hook_when = "just_after_x"
|
|
|
|
stack = stack if remember_stack else (copy.deepcopy(stack) if stack is not None else {})
|
|
tree = parse_expr(Expression) if isinstance(Expression, str) else Expression
|
|
|
|
MAX_HOOK_INDEX = 999
|
|
hit_flags = {"attn1": False, "attn2": False, "dit": False, "attn_unknown": False, "unet": False}
|
|
|
|
def resolve_attn_kind(transformer_options: dict, q: torch.Tensor | None = None, k: torch.Tensor | None = None, v: torch.Tensor | None = None) -> str:
|
|
raw = transformer_options.get(
|
|
"attn_name",
|
|
transformer_options.get(
|
|
"attn_type",
|
|
transformer_options.get(
|
|
"block_attn",
|
|
transformer_options.get("transformer_index", None),
|
|
),
|
|
),
|
|
)
|
|
|
|
if isinstance(raw, str):
|
|
r = raw.lower()
|
|
if r in ("attn1", "self", "self_attn", "self_attention"):
|
|
return "attn1"
|
|
if r in ("attn2", "cross", "cross_attn", "cross_attention"):
|
|
return "attn2"
|
|
return "attn_unknown"
|
|
|
|
if isinstance(raw, (int, float)):
|
|
if int(raw) == 0:
|
|
return "attn1"
|
|
if int(raw) == 1:
|
|
return "attn2"
|
|
|
|
# Fallback heuristika: self-attn má obvykle stejný zdroj pro q/k/v
|
|
if isinstance(q, torch.Tensor) and isinstance(k, torch.Tensor) and isinstance(v, torch.Tensor):
|
|
if q.shape == k.shape == v.shape:
|
|
if (q.data_ptr() == k.data_ptr()) and (k.data_ptr() == v.data_ptr()):
|
|
return "attn1"
|
|
# stejný tvar bývá často pořád self-attn
|
|
return "attn1"
|
|
return "attn2"
|
|
|
|
return "attn_unknown"
|
|
|
|
def resolve_x(total_blocks: int | None) -> int:
|
|
x = int(layer_x)
|
|
if x < 0 and total_blocks is not None:
|
|
x = total_blocks + x
|
|
if total_blocks is not None:
|
|
x = max(0, min(x, total_blocks - 1))
|
|
return x
|
|
|
|
def match_index(idx: int, total_blocks: int | None, stage: str, topts: dict, hook_kind: str) -> bool:
|
|
x = resolve_x(total_blocks)
|
|
return idx == x
|
|
|
|
def target_indices():
|
|
return [int(layer_x)]
|
|
|
|
def side_from_cond_or_uncond(cond_or_uncond):
|
|
if isinstance(cond_or_uncond, (list, tuple)) and len(cond_or_uncond) > 0:
|
|
vals = list(cond_or_uncond)
|
|
uniq = set(vals)
|
|
if len(uniq) == 1:
|
|
idx = int(vals[0])
|
|
if idx == 0:
|
|
return ("positive", 0)
|
|
if idx == 1:
|
|
return ("negative", 1)
|
|
return ("mixed", -1)
|
|
return ("unknown", -1)
|
|
|
|
def run_expr(inp: torch.Tensor, extra_vars: dict | None = None) -> torch.Tensor:
|
|
variables = {
|
|
"inp": inp,
|
|
"sample": inp,
|
|
"hook_kind": "unknown",
|
|
"hook_domain": "unknown",
|
|
"attn_kind": "none",
|
|
"transformer_index": -1.0,
|
|
"is_dit": 0.0,
|
|
"is_unet_block": 0.0,
|
|
"is_attn1": 0.0,
|
|
"is_attn2": 0.0,
|
|
"block_name": "",
|
|
"layer_key": "",
|
|
"layer_id": -1.0,
|
|
"layer": -1.0,
|
|
"i": -1.0,
|
|
"total_blocks": -1.0,
|
|
"has_qkv": 0.0,
|
|
"q": inp,
|
|
"k": inp,
|
|
"v": inp,
|
|
"heads": 0.0,
|
|
"dim_head": 0.0,
|
|
"cond_side": "unknown",
|
|
"cond_index": -1.0,
|
|
"is_positive": 0.0,
|
|
"is_negative": 0.0,
|
|
} | generate_dim_variables(inp)
|
|
|
|
if extra_vars:
|
|
variables.update(extra_vars)
|
|
|
|
for kf, vf in F.items():
|
|
variables[kf] = vf if vf is not None else 0.0
|
|
|
|
visitor = UnifiedMathVisitor(variables, inp.shape, inp.device, state_storage=stack)
|
|
out = visitor.visit(tree)
|
|
return as_tensor(out, inp.shape).to(inp.device)
|
|
|
|
def register_dit(patched_dict, forced_side=None):
|
|
topts = patched_dict["transformer_options"]
|
|
|
|
def dit_wrap(args, extra_args):
|
|
idx = int(args.get("transformer_options", {}).get("block_index", -1))
|
|
total_blocks = args.get("transformer_options", {}).get("total_blocks", None)
|
|
|
|
if not match_index(idx, total_blocks, "dit", args.get("transformer_options", {}), "dit_block"):
|
|
return extra_args["original_block"](args)
|
|
if not hit_flags["dit"]:
|
|
logging.warning(f"[SelectiveGuiderMathNode] DIT HIT idx={idx} total={total_blocks}")
|
|
hit_flags["dit"] = True
|
|
btype = args.get("transformer_options", {}).get("block_type", "dit")
|
|
co = args.get("transformer_options", {}).get("cond_or_uncond", None)
|
|
auto_side, auto_idx = side_from_cond_or_uncond(co)
|
|
side = forced_side if forced_side is not None else auto_side
|
|
idx_side = 0 if side == "positive" else (1 if side == "negative" else auto_idx)
|
|
|
|
meta = {
|
|
"hook_kind": "dit_block",
|
|
"hook_domain": "diffusion",
|
|
"attn_kind": "none",
|
|
"is_dit": 1.0,
|
|
"is_attn1": 0.0,
|
|
"is_attn2": 0.0,
|
|
"layer_id": float(idx),
|
|
"layer": float(idx),
|
|
"i": float(idx),
|
|
"total_blocks": float(total_blocks) if total_blocks is not None else -1.0,
|
|
"block_name": btype,
|
|
"layer_key": f"dit.{btype}.{idx}",
|
|
"has_qkv": 0.0,
|
|
"cond_side": side,
|
|
"cond_index": float(idx_side),
|
|
"is_positive": 1.0 if side == "positive" else 0.0,
|
|
"is_negative": 1.0 if side == "negative" else 0.0,
|
|
}
|
|
|
|
out = extra_args["original_block"](args)
|
|
if isinstance(out, dict) and "img" in out:
|
|
out2 = dict(out)
|
|
out2["img"] = run_expr(out["img"], meta)
|
|
return out2
|
|
return out
|
|
|
|
patches_replace = topts.get("patches_replace", {}).copy()
|
|
dit_map = patches_replace.get("dit", {}).copy()
|
|
for li in target_indices():
|
|
dit_map[("double_block", li)] = dit_wrap
|
|
dit_map[("single_block", li)] = dit_wrap
|
|
patches_replace["dit"] = dit_map
|
|
topts["patches_replace"] = patches_replace
|
|
patched_dict["transformer_options"] = topts
|
|
|
|
def register_attn(patched_dict, forced_side=None):
|
|
topts = patched_dict["transformer_options"].copy()
|
|
|
|
def attn_override(original_attn, q, k, v, heads, **kwargs):
|
|
transformer_options = kwargs.get("transformer_options", {})
|
|
attn_kind = resolve_attn_kind(transformer_options, q=q, k=k, v=v)
|
|
t_index = int(transformer_options.get("transformer_index", -1))
|
|
|
|
if hook_target == "attn1" and attn_kind != "attn1":
|
|
return original_attn(q, k, v, heads, **kwargs)
|
|
if hook_target == "attn2" and attn_kind != "attn2":
|
|
return original_attn(q, k, v, heads, **kwargs)
|
|
|
|
blk = transformer_options.get("block", None)
|
|
if isinstance(blk, (tuple, list)) and len(blk) >= 2:
|
|
stage = str(blk[0])
|
|
layer_id = int(blk[1])
|
|
else:
|
|
stage = "unknown"
|
|
layer_id = int(transformer_options.get("block_index", -1))
|
|
|
|
idx = layer_id
|
|
total_blocks = None
|
|
|
|
if not match_index(idx, total_blocks, stage, transformer_options, attn_kind):
|
|
return original_attn(q, k, v, heads, **kwargs)
|
|
|
|
co = transformer_options.get("cond_or_uncond", None)
|
|
auto_side, auto_idx = side_from_cond_or_uncond(co)
|
|
side = forced_side if forced_side is not None else auto_side
|
|
idx_side = 0 if side == "positive" else (1 if side == "negative" else auto_idx)
|
|
|
|
if not hit_flags.get(attn_kind, False):
|
|
logging.warning(f"[SelectiveGuiderMathNode] {attn_kind.upper()} HIT stage={stage} layer_id={layer_id} side={side}")
|
|
hit_flags[attn_kind] = True
|
|
|
|
meta = {
|
|
"hook_kind": attn_kind,
|
|
"hook_domain": "attention",
|
|
"attn_kind": attn_kind,
|
|
"transformer_index": float(t_index),
|
|
"is_dit": 0.0,
|
|
"is_attn1": 1.0 if attn_kind == "attn1" else 0.0,
|
|
"is_attn2": 1.0 if attn_kind == "attn2" else 0.0,
|
|
"block_name": stage,
|
|
"layer_id": float(layer_id),
|
|
"layer": float(layer_id),
|
|
"i": float(layer_id),
|
|
"total_blocks": -1.0,
|
|
"has_qkv": 1.0,
|
|
"q": q,
|
|
"k": k,
|
|
"v": v,
|
|
"heads": float(heads),
|
|
"dim_head": float(q.shape[-1] // heads) if heads > 0 else 0.0,
|
|
"layer_key": f"{stage}.{layer_id}.{attn_kind}.{t_index}",
|
|
"cond_side": side,
|
|
"cond_index": float(idx_side),
|
|
"is_positive": 1.0 if side == "positive" else 0.0,
|
|
"is_negative": 1.0 if side == "negative" else 0.0,
|
|
}
|
|
|
|
out = original_attn(q, k, v, heads, **kwargs)
|
|
return run_expr(out, meta)
|
|
|
|
topts["optimized_attention_override"] = attn_override
|
|
patched_dict["transformer_options"] = topts
|
|
return patched_dict
|
|
|
|
def register_unet_blocks(patched_dict, forced_side=None):
|
|
topts = patched_dict["transformer_options"].copy()
|
|
patches = topts.get("patches", {}).copy()
|
|
|
|
def _parse_block(transformer_options: dict, fallback_stage: str):
|
|
blk = transformer_options.get("block", None)
|
|
if isinstance(blk, (tuple, list)) and len(blk) >= 2:
|
|
return str(blk[0]), int(blk[1])
|
|
return fallback_stage, int(transformer_options.get("block_index", -1))
|
|
|
|
def _apply_block_expr(t: torch.Tensor, transformer_options: dict, stage: str, idx: int):
|
|
if not match_index(idx, None, stage, transformer_options, "unet_block"):
|
|
return t
|
|
|
|
co = transformer_options.get("cond_or_uncond", None)
|
|
auto_side, auto_idx = side_from_cond_or_uncond(co)
|
|
side = forced_side if forced_side is not None else auto_side
|
|
idx_side = 0 if side == "positive" else (1 if side == "negative" else auto_idx)
|
|
|
|
if not hit_flags["unet"]:
|
|
logging.warning(f"[SelectiveGuiderMathNode] UNET BLOCK HIT stage={stage} layer_id={idx} side={side}")
|
|
hit_flags["unet"] = True
|
|
|
|
meta = {
|
|
"hook_kind": "unet_block",
|
|
"hook_domain": "diffusion",
|
|
"attn_kind": "none",
|
|
"is_dit": 0.0,
|
|
"is_unet_block": 1.0,
|
|
"is_attn1": 0.0,
|
|
"is_attn2": 0.0,
|
|
"block_name": stage,
|
|
"layer_id": float(idx),
|
|
"layer": float(idx),
|
|
"i": float(idx),
|
|
"total_blocks": -1.0,
|
|
"has_qkv": 0.0,
|
|
"layer_key": f"unet.{stage}.{idx}",
|
|
"cond_side": side,
|
|
"cond_index": float(idx_side),
|
|
"is_positive": 1.0 if side == "positive" else 0.0,
|
|
"is_negative": 1.0 if side == "negative" else 0.0,
|
|
}
|
|
return run_expr(t, meta)
|
|
|
|
def input_block_patch(h, transformer_options):
|
|
stage, idx = _parse_block(transformer_options, "input")
|
|
return _apply_block_expr(h, transformer_options, stage, idx)
|
|
|
|
def output_block_patch(h, hsp, transformer_options):
|
|
stage, idx = _parse_block(transformer_options, "output")
|
|
h2 = _apply_block_expr(h, transformer_options, stage, idx)
|
|
return h2, hsp
|
|
|
|
def middle_patch(x, extra_options):
|
|
stage, idx = _parse_block(extra_options, "middle")
|
|
return _apply_block_expr(x, extra_options, stage, idx)
|
|
|
|
patches["input_block_patch"] = list(patches.get("input_block_patch", [])) + [input_block_patch]
|
|
patches["output_block_patch"] = list(patches.get("output_block_patch", [])) + [output_block_patch]
|
|
patches["middle_patch"] = list(patches.get("middle_patch", [])) + [middle_patch]
|
|
|
|
topts["patches"] = patches
|
|
patched_dict["transformer_options"] = topts
|
|
return patched_dict
|
|
|
|
def build_transformers_for_side(side_label=None):
|
|
p = comfy.model_patcher.create_model_options_clone(V.model_options)
|
|
p.setdefault("transformer_options", {})
|
|
if hook_target in ("all", "dit_block"):
|
|
register_dit(p, forced_side=side_label)
|
|
if hook_target in ("all", "unet_block"):
|
|
p = register_unet_blocks(p, forced_side=side_label)
|
|
if hook_target in ("all", "attn1", "attn2"):
|
|
p = register_attn(p, forced_side=side_label)
|
|
return p["transformer_options"]
|
|
|
|
def attach_side_hook(side_name: str):
|
|
if not hasattr(V, "original_conds") or side_name not in V.original_conds:
|
|
return
|
|
tdict = build_transformers_for_side(side_name)
|
|
hook = comfy.hooks.TransformerOptionsHook(
|
|
transformers_dict=tdict,
|
|
hook_scope=comfy.hooks.EnumHookScope.HookedOnly,
|
|
)
|
|
hg = comfy.hooks.HookGroup()
|
|
hg.add(hook)
|
|
|
|
new_conds = []
|
|
for c in V.original_conds[side_name]:
|
|
cc = c.copy()
|
|
old = cc.get("hooks", None)
|
|
cc["hooks"] = comfy.hooks.HookGroup.combine_all_hooks([old, hg])
|
|
new_conds.append(cc)
|
|
V.original_conds[side_name] = new_conds
|
|
|
|
# po parse_expr(...)
|
|
if hasattr(V, "original_conds"):
|
|
if not hasattr(V, "_mrmth_base_original_conds"):
|
|
V._mrmth_base_original_conds = copy.deepcopy(V.original_conds)
|
|
else:
|
|
V.original_conds = copy.deepcopy(V._mrmth_base_original_conds)
|
|
|
|
if hasattr(V, "original_conds"):
|
|
attach_side_hook("positive")
|
|
attach_side_hook("negative")
|
|
logging.warning(f"[SelectiveGuiderMathNode] has original_conds: {list(V.original_conds.keys())}")
|
|
for side in ("positive", "negative"):
|
|
if side in V.original_conds:
|
|
cnt = sum(1 for c in V.original_conds[side] if c.get("hooks", None) is not None)
|
|
logging.warning(f"[SelectiveGuiderMathNode] side={side} conds_with_hooks={cnt}/{len(V.original_conds[side])}")
|
|
else:
|
|
patched = comfy.model_patcher.create_model_options_clone(V.model_options)
|
|
patched.setdefault("transformer_options", {})
|
|
if hook_target in ("all", "dit_block"):
|
|
register_dit(patched, forced_side=None)
|
|
if hook_target in ("all", "unet_block"):
|
|
patched = register_unet_blocks(patched, forced_side=None)
|
|
if hook_target in ("all", "attn1", "attn2"):
|
|
patched = register_attn(patched, forced_side=None)
|
|
V.model_options = patched
|
|
|
|
if hasattr(V, "model_patcher") and V.model_patcher is not None and not getattr(V, "_mrmth_debug_callbacks_added", False):
|
|
def _on_register_all_hook_patches(model_patcher, hooks, target_dict, model_options, registered):
|
|
try:
|
|
reg_len = len(registered) if registered is not None else 0
|
|
in_len = len(hooks) if hooks is not None else 0
|
|
tr_count = len(registered.get_type(comfy.hooks.EnumHookType.TransformerOptions)) if registered is not None else 0
|
|
logging.warning(f"[SelectiveGuiderMathNode] REGISTER hooks_in={in_len} registered={reg_len} transformer_registered={tr_count}")
|
|
if model_options is not None:
|
|
rh = model_options.get("registered_hooks", None)
|
|
logging.warning(f"[SelectiveGuiderMathNode] model_options.registered_hooks={'set' if rh is not None else 'None'}")
|
|
except Exception as e:
|
|
logging.warning(f"[SelectiveGuiderMathNode] REGISTER debug error: {e}")
|
|
|
|
def _on_apply_hooks(model_patcher, hooks):
|
|
try:
|
|
hl = len(hooks) if hooks is not None else 0
|
|
logging.warning(f"[SelectiveGuiderMathNode] APPLY hooks_len={hl}")
|
|
except Exception as e:
|
|
logging.warning(f"[SelectiveGuiderMathNode] APPLY debug error: {e}")
|
|
|
|
V.model_patcher.add_callback_with_key(
|
|
comfy.patcher_extension.CallbacksMP.ON_REGISTER_ALL_HOOK_PATCHES,
|
|
"mrmth_selective_debug",
|
|
_on_register_all_hook_patches,
|
|
)
|
|
V.model_patcher.add_callback_with_key(
|
|
comfy.patcher_extension.CallbacksMP.ON_APPLY_HOOKS,
|
|
"mrmth_selective_debug",
|
|
_on_apply_hooks,
|
|
)
|
|
V._mrmth_debug_callbacks_added = True
|
|
|
|
stack = stack if remember_stack else copy.deepcopy(stack)
|
|
return (V, stack) |