digital-garbageandClaude Haiku 5.5
77748b7a7e
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
616 lines
32 KiB
Python
616 lines
32 KiB
Python
"""SLA: block-sparse attention for MiniMax H3, as a loader choice.
|
|
|
|
Not a module: see common.py's own note -- a plain file at the domain level is a
|
|
helper the loaders import, and ships no node of its own.
|
|
|
|
ComfyUI has no sparse-attention backend for H3, which is why lightx2v's SLA turbo
|
|
LoRA gives no speedup on its own: the LoRA is the *adaptation* to sparsity, not the
|
|
acceleration. This module supplies the missing half — mean-pool Q and a smoothed K
|
|
into blocks, score them with one small matmul, and attend only the top
|
|
``1 - sparsity`` fraction of key blocks per query block. Nothing is trained and
|
|
nothing is loaded; the published SLA files contain only ordinary LoRA tensors.
|
|
|
|
Ported from ComfyUI-H3-SLA-Attention (MIT; refreshed against its 2026-09-21 version),
|
|
whose kernel and block map are vendored from LightX2V (Apache-2.0) — see sla_kernel.py /
|
|
sla_block_map.py for the change lists. What is FunPack's here: this file, the H3 model
|
|
check, and the loader wiring, so turning it on is picking a value in the diffusion model
|
|
loader rather than adding and wiring a node.
|
|
|
|
Two engines run the sparse calls: comfy_kitchen's compiled ``sol_attn`` (int8, the default
|
|
where installed) and the vendored Triton kernel (block size, stabilize_motion and the exact
|
|
Light reference tier are its alone). That pack's Triton int8-QK and tail additions are not
|
|
ported: it marks both untested on hardware, and sol_attn does both natively.
|
|
|
|
The hook is ``transformer_options["optimized_attention_override"]``, which
|
|
``wrap_attn`` consults and H3's one attention call site reaches. The legacy
|
|
``set_model_attn1_patch`` is the SD-UNet path a DiT never consults: a patch installed
|
|
there reports success and silently does nothing, which is why the invocation counter
|
|
below exists and why a run that never sparsified says so in the log.
|
|
"""
|
|
import logging
|
|
|
|
from .._core import dit_hooks as _dit_hooks, patching, traits as _traits
|
|
|
|
SLA_NAME = "sla_h3"
|
|
|
|
_H3_HEAD_DIM = 128
|
|
# Qualified on purpose, per core/traits.py's has_block: "a bare name matches on
|
|
# the class name alone... actively wrong" for a false positive that silently
|
|
# sparsifies a model this was never validated on -- exactly the LTX-collision
|
|
# risk this file's own docstring already warns about, just not guarded against
|
|
# strongly enough before this fix (v4 matched the bare name; found in extensive_
|
|
# testing round 1 on this port). Kept as a local string matching modules/models/
|
|
# minimax_h3's own MODEL_CLASS rather than importing that module -- it pulls in
|
|
# comfy_extras at import time, which this file deliberately does not require
|
|
# just to be importable on a machine with no ComfyUI on the path. Keep the two
|
|
# in sync if the H3 model class is ever renamed.
|
|
_H3_MODEL_CLASS = "comfy.ldm.minimax.model.MiniMaxH3Model"
|
|
|
|
# Validated on an RTX 5090 at 768p/15s (ComfyUI-H3-SLA-Attention's own measurements):
|
|
# 3.7x the attention throughput of stock ComfyUI at sparsity 0.90 / block 64.
|
|
SLA_DEFAULTS = {
|
|
"sparsity_ratio": 0.90,
|
|
"block_size": 64,
|
|
"min_seq_len": 8192,
|
|
"dense_last_steps": 0,
|
|
"protect_audio": True,
|
|
"enabled": True,
|
|
"engine": "comfy_kitchen",
|
|
"dense_steps": "1",
|
|
"method": "sla",
|
|
"tau": 1.3,
|
|
"references": "off",
|
|
"tail": False,
|
|
"stabilize_motion": False,
|
|
}
|
|
ENGINES = ("comfy_kitchen", "triton")
|
|
# sla: each query block keeps the top (1 - sparsity) of key blocks, what SLA turbo LoRAs were distilled against.
|
|
# sol-attn: comfy_kitchen picks per head and block by a threshold (tau), training-free. vsa: FastVideo's cube
|
|
# tiling, for models trained for it -- ComfyUI's own Model Sparse Attention runs it, handed the settings here.
|
|
METHODS = ("sla", "sol-attn", "vsa")
|
|
REFERENCES = ("off", "light", "heavy")
|
|
_LIGHT_KEEP = 0.15 # "light": each reference span keeps its best 15% of blocks for every query
|
|
_CK_BLOCK = 64 # sol_attn's fixed key block
|
|
|
|
try: # ComfyUI with its compiler (PR #16148): a long-lived tensor made mid-graph must be made outside it
|
|
from comfy.model_prefetch import pause_malloc_graph
|
|
except ImportError:
|
|
import contextlib
|
|
pause_malloc_graph = contextlib.nullcontext
|
|
|
|
|
|
def _torch():
|
|
import torch
|
|
return torch
|
|
|
|
|
|
def sla_available():
|
|
"""True when this machine could actually run the kernel.
|
|
|
|
Triton and CUDA, nothing else — the ladder in sla_kernel.py handles differing
|
|
shared-memory limits per architecture. Offered as a choice only where it can run,
|
|
like every other backend in the loader's list.
|
|
"""
|
|
try:
|
|
import torch
|
|
import triton # noqa: F401
|
|
except Exception: # noqa: BLE001
|
|
return False
|
|
return bool(torch.cuda.is_available())
|
|
|
|
|
|
def ck_available():
|
|
"""comfy_kitchen's sol_attn (comfy-kitchen >= 0.2.32) compiled for this GPU."""
|
|
try:
|
|
import comfy_kitchen as ck
|
|
return bool(ck.sol_attn_is_available())
|
|
except Exception: # noqa: BLE001
|
|
return False
|
|
|
|
|
|
def is_h3_model(model):
|
|
"""Whether this patcher holds a MiniMax H3 diffusion model.
|
|
|
|
SLA is only sound on H3: the sparsity ratio it runs at is the one the turbo LoRA
|
|
was distilled to tolerate, and H3's packed [text | cond | audio | video] sequence
|
|
is what the span protection is for. Head shape alone would also match LTX, and
|
|
silently sparsifying LTX attention is a quality loss with no LoRA compensating it
|
|
-- which is also why the class name is matched qualified (module + name), not
|
|
bare: a bare name is exactly as blind to an unrelated same-named class as it is
|
|
to LTX's different head shape.
|
|
|
|
has_block's contract is "does this model CONTAIN one", not "IS this model
|
|
one" -- true today because nothing in v5 wraps diffusion_model in anything
|
|
that could nest an unrelated MiniMaxH3Model submodule (checked: comfy's own
|
|
model_base.py assigns the raw architecture instance directly, and no
|
|
composite/multi-model pipeline exists in this codebase yet). Revisit this
|
|
gate if that ever changes -- the same assumption underpins minimax_h3's own
|
|
equally consequential is_h3() gate, so it would need fixing in both places.
|
|
"""
|
|
return _traits.has_block(model, _H3_MODEL_CLASS)
|
|
|
|
|
|
def block_ranges(spans, blk, nk):
|
|
"""Token spans [(start, stop)] -> merged half-open key-block ranges, rounded outward."""
|
|
out = []
|
|
for a, b in sorted((min(nk, max(0, int(a)) // blk), min(nk, (int(b) + blk - 1) // blk)) for a, b in spans if b > a):
|
|
if a >= b:
|
|
continue
|
|
if out and a <= out[-1][1]:
|
|
out[-1] = (out[-1][0], max(out[-1][1], b))
|
|
else:
|
|
out.append((a, b))
|
|
return out
|
|
|
|
def minus(ranges, cut):
|
|
"""Block ranges with every block in `cut` taken out."""
|
|
for c0, c1 in cut:
|
|
ranges = [p for a, b in ranges for p in ((a, min(b, c0)), (max(a, c1), b)) if p[0] < p[1]]
|
|
return ranges
|
|
|
|
|
|
|
|
def parse_steps(spec):
|
|
""""1,3-4" -> {0, 2, 3}: steps as a person counts them (1 is the first) -> the sampler's 0-based
|
|
indices. A token that is not such a number (0, a letter, past step 1000) is skipped and named."""
|
|
steps, bad = set(), []
|
|
for tok in str(spec or "").replace(" ", "").split(","):
|
|
if not tok or tok == "0": # blank, or "0": no dense steps (the tooltip and the commit say so)
|
|
continue
|
|
a, _, b = tok.partition("-")
|
|
if a.isdigit() and (not b or b.isdigit()) and 1 <= min(int(a), int(b or a)) and max(int(a), int(b or a)) <= 1000:
|
|
lo, hi = sorted((int(a), int(b or a)))
|
|
steps.update(range(lo - 1, hi))
|
|
else:
|
|
bad.append(tok)
|
|
return frozenset(steps), bad
|
|
|
|
|
|
def _runs(numbers):
|
|
"""[1, 2, 3, 6] -> "1-3, 6"."""
|
|
out = []
|
|
for n in sorted(numbers):
|
|
if out and n == out[-1][1] + 1:
|
|
out[-1][1] = n
|
|
else:
|
|
out.append([n, n])
|
|
return ", ".join(str(a) if a == b else f"{a}-{b}" for a, b in out)
|
|
|
|
|
|
def spans(payload):
|
|
"""H3's packed layout -> (video start, spans to keep exact, visual reference spans).
|
|
|
|
Exact: the language tokens and every audio stream (target, reference, conditioning) --
|
|
audio is ~1% of the sequence, so plain top-k routinely drops all of it and the
|
|
soundtrack degrades while the picture looks fine. Visual references (the vision tokens
|
|
inside the text span, conditioning frames, reference images) stay sparse unless
|
|
`references` asks otherwise. No layout: nothing is protected, rather than guessed at.
|
|
"""
|
|
layout = payload.get("layout") if payload else None
|
|
tags = payload.get("text_token_tags") if payload else None
|
|
video, keep, refs = 0, [], []
|
|
for seg in getattr(layout, "segments", ()) or ():
|
|
if len(seg) != 3:
|
|
continue
|
|
a, b, kind = int(seg[0]), int(seg[1]), seg[2]
|
|
if kind == "text":
|
|
try:
|
|
t = [int(x) for x in (tags.reshape(-1).tolist() if hasattr(tags, "reshape") else tags)]
|
|
except TypeError:
|
|
t = None
|
|
if t is None or len(t) != b - a: # untagged: all of it is language
|
|
keep.append((a, b))
|
|
continue
|
|
i = 0
|
|
for j in range(1, len(t) + 1): # tag runs: 0 = a vision token, else language
|
|
if j == len(t) or t[j] != t[i]:
|
|
(refs if t[i] == 0 else keep).append((a + i, a + j))
|
|
i = j
|
|
elif kind in ("audio", "ref_audio", "cond_audio"):
|
|
keep.append((a, b))
|
|
elif kind in ("cond", "ref_img"):
|
|
refs.append((a, b))
|
|
elif kind == "video" and not video:
|
|
video = a
|
|
return video, keep, refs
|
|
|
|
|
|
def new_state():
|
|
return {
|
|
"calls": 0, # sparse invocations this run
|
|
"dense": 0, # fall-throughs this run
|
|
"step": 0,
|
|
"seq": 0,
|
|
"kept": 0,
|
|
"blocks": 0,
|
|
"pinned": 0,
|
|
"backend": None, # what we displaced
|
|
"failed": None, # first kernel failure, if any
|
|
"layer": 0, # this step's call count: the Nth call is the same layer every step
|
|
"history": {}, # layer -> last step's choices, for stabilize_motion
|
|
"order": None, # comfy_kitchen: (S, protected spans, key order, sink blocks), same every layer
|
|
"last": None, # last step index read from the schedule
|
|
"dense_at": set(), # steps (counted from 1) that ran at full attention this run
|
|
"n": 0, # this run's step count
|
|
"summed": False, # this run's summary is logged
|
|
}
|
|
|
|
|
|
def _summarise(state, cfg):
|
|
"""One line per sampling run. Never one per block — there are 50 of those."""
|
|
if state["calls"] == 0:
|
|
logging.warning(
|
|
"[FunPack SLA] installed but never invoked — attention was NOT sparsified "
|
|
"(%d dense fall-throughs). Check that the MODEL reaching the sampler is the "
|
|
"one this loader produced.", state["dense"])
|
|
return
|
|
real = 1.0 - (state["kept"] / state["blocks"]) if state["blocks"] else 0.0
|
|
logging.info(
|
|
"[FunPack SLA] %s | %d calls | S=%d | blocks %d/%d kept (%.1f%% sparse, asked %.0f%%) "
|
|
"| %d pinned | %d dense fall-throughs | displaced %s",
|
|
cfg["engine"], state["calls"], state["seq"], state["kept"], state["blocks"], real * 100.0,
|
|
cfg["sparsity_ratio"] * 100.0, state["pinned"], state["dense"], state["backend"] or "?")
|
|
logging.info("[FunPack SLA] full attention on %s of %d", f"steps {_runs(state['dense_at'])}"
|
|
if state["dense_at"] else "no step", state["n"])
|
|
if state["failed"] is not None:
|
|
logging.warning("[FunPack SLA] kernel fell back to dense at least once: %s",
|
|
state["failed"])
|
|
|
|
|
|
def _mean_pool_torch(x, blk):
|
|
"""``(B, L, H, D)`` -> ``(B, H, ceil(L/blk), D)`` fp32, the mean over each block: the same numbers as the
|
|
Triton pool in sla_block_map, in plain torch, so Light works without Triton."""
|
|
import torch
|
|
B, L, H, D = x.shape
|
|
pad = (-L) % blk
|
|
x = torch.nn.functional.pad(x.float().transpose(1, 2), (0, 0, 0, pad)) # (B, H, L+pad, D)
|
|
counts = torch.nn.functional.pad(torch.ones(L, device=x.device), (0, pad)).reshape(1, 1, -1, blk, 1).sum(dim=3)
|
|
return x.reshape(B, H, -1, blk, D).sum(dim=3) / counts # the last, short block averages its real rows only
|
|
|
|
|
|
def _ck_attention(state, cfg, qb, kb, vb, keep, refs, scale):
|
|
"""comfy_kitchen's sol_attn: one 64-token key-block range (`sink_blocks`) is always exact. Attention
|
|
does not care about key order, so when the protected blocks are not one run, K and V are put in an
|
|
order where they are -- a copy per call, needed only then. -> (output, blocks, pinned)."""
|
|
import comfy_kitchen as ck
|
|
|
|
torch = _torch()
|
|
S = kb.shape[1]
|
|
nk = (S + _CK_BLOCK - 1) // _CK_BLOCK
|
|
light = cfg["references"] == "light"
|
|
cached = state["order"]
|
|
if cached and cached[0] == S and cached[1] == keep and not light:
|
|
_, _, idx, sink = cached
|
|
else:
|
|
pinned = block_ranges(keep, _CK_BLOCK, nk)
|
|
if light: # one set for every query here: ranked by the average query, not each one (Triton does that)
|
|
score = (_mean_pool_torch(qb, _CK_BLOCK).mean(dim=2, keepdim=True)
|
|
@ (_mean_pool_torch(kb, _CK_BLOCK) - kb.mean(dim=1, dtype=torch.float32)[:, :, None, :]).transpose(-1, -2))
|
|
score = score.mean(dim=(0, 1, 2))
|
|
for a, b in minus(block_ranges(refs, _CK_BLOCK, nk), pinned):
|
|
best = torch.topk(score[a:b], max(1, round(_LIGHT_KEEP * (b - a)))).indices + a
|
|
pinned += [(i, i + 1) for i in best.tolist()]
|
|
pinned = block_ranges([(a * _CK_BLOCK, b * _CK_BLOCK) for a, b in pinned], _CK_BLOCK, nk)
|
|
if len(pinned) <= 1:
|
|
idx, sink = None, list(pinned[0]) if pinned else [0, 0]
|
|
else:
|
|
first = [i for a, b in pinned for i in range(a, b)]
|
|
taken = set(first)
|
|
blocks = first + [i for i in range(nk) if i not in taken]
|
|
with pause_malloc_graph():
|
|
idx = torch.cat([torch.arange(i * _CK_BLOCK, min(S, (i + 1) * _CK_BLOCK), device=kb.device)
|
|
for i in blocks])
|
|
sink = [0, len(first)]
|
|
if not light:
|
|
state["order"] = (S, keep, idx, sink)
|
|
if idx is not None:
|
|
kb, vb = kb.index_select(1, idx), vb.index_select(1, idx)
|
|
cast = qb.dtype != torch.bfloat16 # the compiled kernel takes bf16 only
|
|
q, k, v = ((t.to(torch.bfloat16) if cast else t) for t in (qb, kb, vb))
|
|
if cfg["method"] == "sol-attn": # threshold routing: the pooled tail is part of the method
|
|
out = ck.sol_attn(q, k, v, tau=cfg["tau"], scale=scale, sink_blocks=sink, sink_q=[0, 0])
|
|
else:
|
|
out = ck.sol_attn(q, k, v, scale=scale, sink_blocks=sink, sink_q=[0, 0],
|
|
topk_ratio=1.0 - cfg["sparsity_ratio"], tail=cfg["tail"])
|
|
return (out.to(qb.dtype) if cast else out), nk, sink[1] - sink[0]
|
|
|
|
|
|
def make_override(state, cfg, dense_fn=None, dense_label=None):
|
|
"""The `optimized_attention_override` wrap_attn hands every attention call to.
|
|
|
|
`cfg` is install_sla's resolved settings. `dense_fn` is another override (same
|
|
signature) handling everything SLA does not: the text refiner, masked calls, the dense
|
|
steps, any non-H3 model. There is only one override slot, so without this, choosing SLA
|
|
would silently discard the backend the user picked and drop those calls onto whatever
|
|
ComfyUI was launched with. Sparse where sparsity applies, the chosen backend everywhere else.
|
|
"""
|
|
torch = _torch()
|
|
ok_dtypes = (torch.bfloat16, torch.float16)
|
|
topk_ratio = 1.0 - cfg["sparsity_ratio"]
|
|
blkq = cfg["block_size"]
|
|
# BLKK=64 is not a typo. On sm_120 the 128x128 tile needs 160 KB of shared memory
|
|
# against a ~99 KB limit and cannot launch at all; 128x64 both fits and measured
|
|
# fastest. LightX2V picks the same split for its sage2 path off sm90.
|
|
blkk = 64 if blkq == 128 else blkq
|
|
heavy = cfg["references"] == "heavy"
|
|
ref_keep = _LIGHT_KEEP if cfg["references"] == "light" else None
|
|
|
|
def override(func, q, k, v, heads, mask=None, attn_precision=None,
|
|
skip_reshape=False, skip_output_reshape=False, **kwargs):
|
|
def dense():
|
|
state["dense"] += 1
|
|
run = (lambda *a, **kw: dense_fn(func, *a, **kw)) if dense_fn is not None else func
|
|
return run(q, k, v, heads, mask=mask, attn_precision=attn_precision,
|
|
skip_reshape=skip_reshape,
|
|
skip_output_reshape=skip_output_reshape, **kwargs)
|
|
|
|
if state["backend"] is None:
|
|
state["backend"] = dense_label or getattr(func, "__name__", repr(func))
|
|
|
|
to = kwargs.get("transformer_options") or {}
|
|
|
|
# Anything that is not the packed H3 self-attention goes straight through. The
|
|
# min_seq_len guard is what keeps the 2-block token refiner (S = text length)
|
|
# and low-resolution runs dense, where selection costs more than it saves.
|
|
if (
|
|
not skip_reshape
|
|
or mask is not None
|
|
or q.ndim != 4
|
|
or q.shape[-1] != _H3_HEAD_DIM
|
|
or q.dtype not in ok_dtypes
|
|
or q.shape[2] < cfg["min_seq_len"]
|
|
or to.get("_funpack_sla_dense", False)
|
|
):
|
|
return dense()
|
|
|
|
try:
|
|
B, H, S, D = q.shape
|
|
layer = state["layer"]
|
|
state["layer"] = layer + 1
|
|
|
|
# [1, H, S, D] -> [1, S, H, D]. H3 builds q/k/v as [S, H, D] and transposes
|
|
# for the call, so this transposes back onto the original memory and the
|
|
# copy is a no-op. A BHSD kernel would cost a real ~1.3 GB copy per tensor.
|
|
# Each checked on its own: newer ComfyUI hands K over in another layout than Q (its fused
|
|
# norm+RoPE), and assuming they match failed every call into dense.
|
|
qb, kb, vb = (t if t.is_contiguous() else t.contiguous() for t in (x.transpose(1, 2) for x in (q, k, v)))
|
|
|
|
video, keep, refs = to.get("_funpack_sla_spans") or (0, (), ())
|
|
keep = tuple(keep if cfg["protect_audio"] else ()) + tuple(refs if heavy else ())
|
|
|
|
if cfg["engine"] == "comfy_kitchen":
|
|
out, blocks, pinned = _ck_attention(state, cfg, qb, kb, vb, keep, refs, D ** -0.5)
|
|
kept = max(1, round(topk_ratio * blocks)) + pinned
|
|
else:
|
|
try:
|
|
from .sla_block_map import get_block_map
|
|
from .sla_kernel import block_sparse_attention
|
|
except ImportError:
|
|
from sla_block_map import get_block_map
|
|
from sla_kernel import block_sparse_attention
|
|
stable = cfg["stabilize_motion"]
|
|
lut, kept, history = get_block_map(
|
|
qb, kb, topk_ratio, blkq, blkk, protect=keep, refs=() if heavy else refs,
|
|
ref_keep=ref_keep, prev=state["history"].get(layer) if stable else None,
|
|
sticky_from=video, remember=stable)
|
|
if stable: # one buffer per layer, refilled in place: what a captured graph can replay
|
|
held = state["history"].get(layer)
|
|
if held is None or held.shape != history.shape:
|
|
with pause_malloc_graph():
|
|
held = state["history"][layer] = torch.empty_like(history)
|
|
held.copy_(history)
|
|
out = block_sparse_attention(qb, kb, vb, lut, kept, blkq, blkk)
|
|
blocks = (S + blkk - 1) // blkk
|
|
pinned = sum(b - a for a, b in block_ranges(keep, blkk, blocks))
|
|
|
|
state["calls"] += 1
|
|
state["seq"] = S
|
|
state["kept"] = min(kept, blocks)
|
|
state["blocks"] = blocks
|
|
state["pinned"] = pinned
|
|
|
|
if skip_output_reshape:
|
|
return out.transpose(1, 2)
|
|
return out.reshape(B, S, H * D)
|
|
|
|
except Exception as exc: # noqa: BLE001 - a bad kernel must not kill the run
|
|
if state["failed"] is None:
|
|
state["failed"] = "%s: %s" % (exc.__class__.__name__, exc)
|
|
# Warning, not debug: a silent fall to dense is close enough to sparse speed
|
|
# to go unnoticed from timing alone.
|
|
logging.warning("[FunPack SLA] %s kernel failed, dense from here this run: %s",
|
|
cfg["engine"], state["failed"], exc_info=True)
|
|
return dense()
|
|
|
|
return override
|
|
|
|
|
|
def make_wrapper(state, cfg):
|
|
"""DIFFUSION_MODEL wrapper: per-step state, and the end-of-run summary.
|
|
|
|
Registered once and then reused — ComfyUI caches node outputs, so this closure
|
|
outlives a single sampling run. The step counter therefore has to reset itself, or
|
|
every run after the first drifts permanently into the trailing-dense window and
|
|
silently stops sparsifying.
|
|
"""
|
|
dense_steps = cfg["dense_steps"]
|
|
last = cfg["dense_last_steps"]
|
|
|
|
def wrapper(executor, x, timestep, context, transformer_options={},
|
|
minimax_payload=None, **kwargs):
|
|
to = transformer_options
|
|
new_run = lambda: (state.update(step=0, calls=0, dense=0, failed=None, order=None, last=None, summed=False),
|
|
state["history"].clear(), state["dense_at"].clear())
|
|
counted = not _dit_hooks.probing(to)
|
|
if to.get("sigmas") is not None and to.get("sample_sigmas") is not None:
|
|
# The schedule says which step this is, so a cancelled run or a sampler calling the model twice
|
|
# a step (heun, res_2s) cannot shift the dense steps. A call between scheduled sigmas (a
|
|
# sampler's second evaluation) belongs to the step it is inside.
|
|
where = _dit_hooks.current_step(to)
|
|
n_steps = where[1] if where else max(1, len(to["sample_sigmas"]) - 1)
|
|
if where and counted:
|
|
idx = where[0]
|
|
if state["last"] is not None and (idx < state["last"] or (state["summed"] and idx <= state["last"])):
|
|
new_run()
|
|
state["last"], state["step"] = idx, idx + 1
|
|
else: # no schedule to read (a caller that is not ComfyUI's sampler): count calls
|
|
n_steps = max(1, len(to.get("sample_sigmas", [])) - 1)
|
|
# A throwaway call (a probe, late-branch's weakened copy) is not a step of the
|
|
# schedule: counting it would slide the dense steps onto the wrong call.
|
|
if counted and state["step"] >= n_steps:
|
|
new_run()
|
|
if counted:
|
|
state["step"] += 1
|
|
state["layer"] = 0
|
|
|
|
# The layout lives on the payload, which never reaches the attention call site,
|
|
# so the wrapper is the only place it can be picked up.
|
|
video, keep, refs = spans(minimax_payload)
|
|
to["_funpack_sla_spans"] = (video, tuple(keep), tuple(refs))
|
|
to["_funpack_sla_dense"] = bool(
|
|
(last > 0 and state["step"] > n_steps - last) or (state["step"] - 1) in dense_steps)
|
|
state["n"] = n_steps
|
|
if to["_funpack_sla_dense"] and counted:
|
|
state["dense_at"].add(state["step"])
|
|
|
|
# Forward minimax_payload only when H3 actually supplied one: every other
|
|
# diffusion model would raise TypeError on the unexpected kwarg, turning a
|
|
# graceful no-op into a crash mid-sampling.
|
|
if minimax_payload is not None:
|
|
kwargs["minimax_payload"] = minimax_payload
|
|
try:
|
|
# The executor itself, not executor.original: that skips every wrapper added after this one.
|
|
out = executor(x, timestep, context, transformer_options=transformer_options, **kwargs)
|
|
except Exception:
|
|
state["history"].clear() # an OOM must not leave the next attempt less memory
|
|
raise
|
|
|
|
if counted and state["step"] >= n_steps and not state["summed"]:
|
|
_summarise(state, cfg)
|
|
state["summed"] = True
|
|
state["history"].clear() # only needed step to step, not while the model sits cached
|
|
return out
|
|
|
|
return wrapper
|
|
|
|
|
|
def install_sla(model, sparsity_ratio=None, block_size=None, min_seq_len=None,
|
|
dense_last_steps=None, protect_audio=None, enabled=None,
|
|
dense_fn=None, dense_label=None, engine=None, dense_steps=None,
|
|
references=None, tail=None, stabilize_motion=None, method=None, tau=None):
|
|
"""Give `model` block-sparse H3 attention. Returns (model, status line, installed).
|
|
|
|
Weights are untouched: this installs an attention override and a per-step wrapper on
|
|
a clone. A model that is not MiniMax H3 comes back unchanged and says so — the
|
|
sparsity is only safe at the ratio H3's turbo LoRA was distilled to tolerate.
|
|
|
|
`installed` is False whenever SLA did not take, so the caller can fall back to
|
|
installing the chosen backend on its own rather than leaving the model with nothing.
|
|
"""
|
|
given = dict(sparsity_ratio=sparsity_ratio, block_size=block_size, min_seq_len=min_seq_len,
|
|
dense_last_steps=dense_last_steps, protect_audio=protect_audio, enabled=enabled,
|
|
engine=engine, dense_steps=dense_steps, references=references, tail=tail,
|
|
stabilize_motion=stabilize_motion, method=method, tau=tau)
|
|
cfg = {key: SLA_DEFAULTS[key] if value is None else value for key, value in given.items()}
|
|
|
|
if not bool(cfg["enabled"]):
|
|
# A dense baseline without touching the backend choice, so an A/B is one click
|
|
# and the settings you were testing are still on the node when you switch back.
|
|
return model, "SLA: off (dense baseline)", False
|
|
if not is_h3_model(model):
|
|
return model, ("SLA: skipped — not a MiniMax H3 model"), False
|
|
triton_ok, ck_ok = sla_available(), ck_available()
|
|
if not (triton_ok or ck_ok):
|
|
return model, ("SLA: skipped — this machine has neither comfy_kitchen's sol_attn nor CUDA+Triton"), False
|
|
|
|
notes = []
|
|
if cfg["method"] not in METHODS:
|
|
cfg["method"] = SLA_DEFAULTS["method"]
|
|
if cfg["method"] == "vsa":
|
|
return _install_vsa(model, cfg, dense_fn, dense_label)
|
|
cfg.update(sparsity_ratio=float(cfg["sparsity_ratio"]), block_size=int(cfg["block_size"]),
|
|
min_seq_len=int(cfg["min_seq_len"]), dense_last_steps=int(cfg["dense_last_steps"]),
|
|
protect_audio=bool(cfg["protect_audio"]), tail=bool(cfg["tail"]),
|
|
stabilize_motion=bool(cfg["stabilize_motion"]),
|
|
references=cfg["references"] if cfg["references"] in REFERENCES else "off")
|
|
if cfg["engine"] not in ENGINES:
|
|
cfg["engine"] = SLA_DEFAULTS["engine"]
|
|
if cfg["engine"] == "comfy_kitchen" and not ck_ok:
|
|
cfg["engine"] = "triton"
|
|
notes.append("comfy_kitchen's sol_attn is not here (needs comfy-kitchen >= 0.2.32): the Triton kernel runs")
|
|
elif cfg["engine"] == "triton" and not triton_ok:
|
|
cfg["engine"] = "comfy_kitchen"
|
|
notes.append("no Triton here: comfy_kitchen's sol_attn runs")
|
|
if cfg["method"] == "sol-attn" and cfg["engine"] != "comfy_kitchen":
|
|
if ck_ok:
|
|
cfg["engine"] = "comfy_kitchen"
|
|
notes.append("sol-attn is comfy_kitchen's: that engine runs")
|
|
else:
|
|
cfg["method"] = "sla"
|
|
notes.append("sol-attn needs comfy_kitchen's sol_attn (comfy-kitchen 0.2.32+), not here: sla runs instead")
|
|
cfg["tau"] = float(cfg["tau"])
|
|
cfg["dense_steps"], bad = parse_steps(cfg["dense_steps"])
|
|
if bad:
|
|
notes.append(f"dense_steps: ignored {', '.join(bad)} (steps count from 1: '1' is the first, '1-2' the first two)")
|
|
if cfg["engine"] == "comfy_kitchen":
|
|
ignored = [n for n, on in (("block_size", cfg["block_size"] != 64), ("stabilize_motion", cfg["stabilize_motion"])) if on]
|
|
if ignored:
|
|
notes.append(f"{', '.join(ignored)}: Triton engine only, ignored (sol_attn routes 64-token blocks, per step)")
|
|
elif cfg["tail"]:
|
|
notes.append("tail: comfy_kitchen engine only, ignored")
|
|
|
|
state = new_state()
|
|
patched = patching.clone(model)
|
|
to = patched.model_options.get("transformer_options", {}).copy()
|
|
to["optimized_attention_override"] = make_override(state, cfg, dense_fn=dense_fn, dense_label=dense_label)
|
|
patched.model_options["transformer_options"] = to
|
|
patched.add_wrapper_with_key("diffusion_model", "funpack_sla_state", make_wrapper(state, cfg))
|
|
|
|
blk = cfg["block_size"] if cfg["engine"] == "triton" else 64
|
|
how = f"tau={cfg['tau']:.2f}" if cfg["method"] == "sol-attn" else f"sparsity={cfg['sparsity_ratio']:.2f}"
|
|
return patched, (f"SLA on | {cfg['method']} on {cfg['engine']} | {how} BLK={blk} "
|
|
f"min_seq_len={cfg['min_seq_len']} dense_last_steps={cfg['dense_last_steps']} "
|
|
f"full attention on steps {_runs(i + 1 for i in cfg['dense_steps']) or 'none'} and the last {cfg['dense_last_steps']} protect_audio={cfg['protect_audio']} "
|
|
f"references={cfg['references']} tail={cfg['tail']} stabilize_motion={cfg['stabilize_motion']} "
|
|
f"| dense calls -> {dense_label or 'as launched'}"
|
|
+ "".join(f"\nSLA: {n}" for n in notes)), True
|
|
|
|
|
|
def _install_vsa(model, cfg, dense_fn, dense_label):
|
|
"""VSA through ComfyUI's own Model Sparse Attention (cube tiling and the trained coarse branch are
|
|
its), the chosen backend underneath for every dense call. Its own sigma window keeps the first
|
|
20% of the schedule dense; dense_steps and the span protection are this file's and do not apply."""
|
|
try:
|
|
from comfy_extras.nodes_sparse_attention import apply_block_sparse_attention
|
|
except ImportError:
|
|
return model, "SLA: vsa needs ComfyUI's Model Sparse Attention (ComfyUI 0.39 or later), not here", False
|
|
if not ck_available():
|
|
return model, "SLA: vsa needs comfy_kitchen's sol_attn (comfy-kitchen 0.2.32+), not here", False
|
|
base = patching.clone(model)
|
|
if dense_fn is not None: # Model Sparse Attention hands what it does not take to the override it found
|
|
to = base.model_options.get("transformer_options", {}).copy()
|
|
to["optimized_attention_override"] = dense_fn
|
|
base.model_options["transformer_options"] = to
|
|
keep = 1.0 - float(cfg["sparsity_ratio"])
|
|
try:
|
|
patched = apply_block_sparse_attention(
|
|
base, tau=float(cfg["tau"]), topk_ratio=keep, vsa=True, start_percent=0.2, end_percent=1.0,
|
|
min_tokens=int(cfg["min_seq_len"]), dense_blocks=set(), sink_conditioning="exact_kv_and_rows",
|
|
extra_tokens=0, verbose=False)
|
|
except (TypeError, ValueError) as exc: # another ComfyUI's version of it, or a model it refuses
|
|
return model, f"SLA: vsa could not be set up by this ComfyUI's Model Sparse Attention: {exc}", False
|
|
notes = [f"SLA on | vsa (ComfyUI's Model Sparse Attention) | keeps {keep:.0%} of video cubes, first 20% of "
|
|
f"the schedule dense, min_seq_len={int(cfg['min_seq_len'])} | dense calls -> {dense_label or 'as launched'}"]
|
|
try:
|
|
trained = getattr(patched.get_model_object("diffusion_model").blocks[0].attn, "to_gate_compress", None) is not None
|
|
except Exception: # noqa: BLE001 - a description must never fail the load
|
|
trained = True
|
|
if not trained:
|
|
notes.append("SLA: this model has no VSA layers (to_gate_compress): it was not trained for VSA, so it runs "
|
|
"without VSA's coarse branch and quality may drop. sla or sol-attn suits it better.")
|
|
ignored = [n for n, on in (("dense_steps", cfg["dense_steps"].strip() not in ("", "1")), ("references", cfg["references"] != "off"),
|
|
("stabilize_motion", cfg["stabilize_motion"]), ("tail", cfg["tail"])) if on]
|
|
if ignored:
|
|
notes.append(f"SLA: {', '.join(ignored)}: not used by vsa")
|
|
return patched, "\n".join(notes), True
|