File size: 9,290 Bytes

b54cde7

# SPDX-License-Identifier: MIT
"""
TurboQuant ComfyUI Nodes

Two nodes:
  1. TurboQuantPatch  - Experimental attention K/V recompression patch
  2. TurboQuantInfo   - Shows observed compression stats for a patched model
"""

import torch
import math
from typing import Optional
try:
    from . import tq3_core
except ImportError:  # imported as a top-level module rather than as a package
    import tq3_core


class TQ3KVCacheWrapper:
    """
    Wraps a standard KV cache to transparently compress/decompress
    using TQ3 quantization.

    Stored tensors are kept in TQ3 format (dict of norms, scales, packed).
    On read, they are decompressed back to float/half.
    """

    def __init__(self, enabled=True):
        self.enabled = enabled
        self._cache = {}  # key -> tq3 dict or raw tensor
        self._stats = {
            "compressed_bytes": 0,
            "original_bytes": 0,
            "num_stores": 0,
        }

    def store(self, key: str, tensor: torch.Tensor) -> None:
        """Compress and store a KV tensor."""
        if not self.enabled:
            self._cache[key] = tensor
            return

        orig_dtype = tensor.dtype
        orig_device = tensor.device
        D = tensor.shape[-1]

        # TQ3 requires last dim divisible by 128
        if D % tq3_core.TQ3_BLOCK != 0:
            # Pad to next multiple of 128
            pad_to = math.ceil(D / tq3_core.TQ3_BLOCK) * tq3_core.TQ3_BLOCK
            pad_size = pad_to - D
            tensor_padded = torch.nn.functional.pad(tensor.float(), (0, pad_size))
        else:
            tensor_padded = tensor.float()
            pad_size = 0

        tq = tq3_core.tq3_quantize(tensor_padded)
        tq["_orig_dtype"] = orig_dtype
        tq["_orig_device"] = orig_device
        tq["_pad_size"] = pad_size

        self._cache[key] = tq

        # Track stats
        compressed_bytes = (
            tq["packed"].numel() * tq["packed"].element_size()
            + tq["norms"].numel() * tq["norms"].element_size()
            + tq["scales"].numel() * tq["scales"].element_size()
        )
        original_bytes = tensor.numel() * tensor.element_size()
        self._stats["compressed_bytes"] += compressed_bytes
        self._stats["original_bytes"] += original_bytes
        self._stats["num_stores"] += 1

    def load(self, key: str) -> Optional[torch.Tensor]:
        """Decompress and return a KV tensor."""
        entry = self._cache.get(key)
        if entry is None:
            return None

        if isinstance(entry, torch.Tensor):
            return entry

        # TQ3 dict
        result = tq3_core.tq3_dequantize(entry)
        pad_size = entry.get("_pad_size", 0)
        if pad_size > 0:
            result = result[..., :-pad_size]

        orig_dtype = entry.get("_orig_dtype", torch.float16)
        orig_device = entry.get("_orig_device", "cpu")
        return result.to(dtype=orig_dtype, device=orig_device)

    def clear(self):
        self._cache.clear()
        self._stats = {"compressed_bytes": 0, "original_bytes": 0, "num_stores": 0}

    @property
    def compression_ratio(self) -> float:
        if self._stats["original_bytes"] == 0:
            return 0.0
        return self._stats["original_bytes"] / max(self._stats["compressed_bytes"], 1)

    @property
    def savings_mb(self) -> float:
        diff = self._stats["original_bytes"] - self._stats["compressed_bytes"]
        return diff / (1024 * 1024)

    def stats_string(self) -> str:
        ratio = self.compression_ratio
        orig_mb = self._stats["original_bytes"] / (1024 * 1024)
        comp_mb = self._stats["compressed_bytes"] / (1024 * 1024)
        return (
            f"TurboQuant TQ3 KV Cache Stats\n"
            f"-----------------------------\n"
            f"Stores:      {self._stats['num_stores']}\n"
            f"Original:    {orig_mb:.1f} MB\n"
            f"TQ3 size:    {comp_mb:.1f} MB\n"
            f"Ratio:       {ratio:.2f}x\n"
            f"Savings:     {self.savings_mb:.1f} MB"
        )


# Global wrapper instance so the info node can read stats
_active_wrapper = None


def _make_attn_patch(wrapper: TQ3KVCacheWrapper):
    """
    Create an attention patch function compatible with ComfyUI's
    model_patcher system.

    ComfyUI patches are callables that receive (q, k, v, extra_options)
    and return (q, k, v). We intercept to compress/decompress KV.

    For the KV cache compression use case, the patch:
    1. On store (when new K,V are computed): compress and cache
    2. On load (when cached K,V are reused): decompress

    In diffusion models, attention is recomputed each step, so the
    main benefit is reducing the *peak* memory of the KV tensors
    that coexist during the attention computation. The patch quantizes
    K and V *in transit* so they occupy less VRAM while stored.
    """

    step_counter = [0]

    def attn_patch(q, k, v, extra_options):
        nonlocal step_counter

        # Compress K and V to TQ3, then immediately decompress.
        # This forces the memory layout through the compressed format,
        # meaning the original FP16 K/V can be freed sooner by PyTorch's
        # allocator. The decompressed versions are fresh allocations
        # that only live for the attention matmul.

        step_counter[0] += 1
        step_id = step_counter[0]

        # ComfyUI coerces a None context away before calling an attn1 patch
        # (attention.py: "if context_attn1 is None: context_attn1 = n") but the
        # attn2 path has no such guard, so k/v arrive as None whenever the block
        # runs without conditioning. Pass those through untouched.
        if k is None or v is None:
            return q, k, v

        # Only compress if dim is >= 128 (otherwise overhead isn't worth it)
        if k.shape[-1] >= tq3_core.TQ3_BLOCK:
            k_key = f"k_{step_id}"
            v_key = f"v_{step_id}"
            wrapper.store(k_key, k)
            wrapper.store(v_key, v)

            # Decompress - original k,v can now be freed
            k = wrapper.load(k_key)
            v = wrapper.load(v_key)

            # Clean up compressed copies (they served their purpose)
            if k_key in wrapper._cache:
                del wrapper._cache[k_key]
            if v_key in wrapper._cache:
                del wrapper._cache[v_key]

        return q, k, v

    return attn_patch


class TurboQuantPatch:
    """
    Experimentally round-trips attention K/V tensors through TQ3 during attention.

    This does not yet replace ComfyUI's persistent KV cache. It is primarily
    useful for validating TQ3 quality and measuring the observed compression
    characteristics of the intermediate K/V tensors.
    """

    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "model": ("MODEL",),
                "enabled": ("BOOLEAN", {"default": True}),
            },
        }

    RETURN_TYPES = ("MODEL",)
    RETURN_NAMES = ("model",)
    FUNCTION = "patch"
    CATEGORY = "TurboQuant"
    DESCRIPTION = (
        "Experimentally round-trips model attention K/V tensors through TQ3. "
        "Useful for quality and compression experiments; not a persistent KV cache yet."
    )

    def patch(self, model, enabled):
        global _active_wrapper

        if not enabled:
            _active_wrapper = None
            return (model,)

        wrapper = TQ3KVCacheWrapper(enabled=True)
        _active_wrapper = wrapper

        # Clone the model patcher so we don't modify the original
        patched = model.clone()

        # Register the attention patch via ComfyUI's patcher API
        # "attn1" = self-attention, "attn2" = cross-attention
        attn_patch = _make_attn_patch(wrapper)
        patched.set_model_attn1_patch(attn_patch)
        patched.set_model_attn2_patch(attn_patch)

        return (patched,)


class TurboQuantInfo:
    """
    Displays observed TurboQuant compression statistics.

    Connect after TurboQuantPatch to see the cumulative compression ratio
    and byte savings observed during generation.
    """

    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "model": ("MODEL",),
            },
        }

    RETURN_TYPES = ("STRING",)
    RETURN_NAMES = ("stats",)
    FUNCTION = "info"
    CATEGORY = "TurboQuant"
    OUTPUT_NODE = True
    DESCRIPTION = "Shows observed TQ3 compression ratio and byte savings."

    def info(self, model):
        global _active_wrapper

        if _active_wrapper is not None:
            stats = _active_wrapper.stats_string()
        else:
            # Estimate based on model config
            stats = (
                "TurboQuant TQ3 KV Cache\n"
                "-----------------------\n"
                "Status: Not yet active (no inference run)\n"
                "Expected compression: ~4.5x\n"
                "Encoding: 3-bit Lloyd-Max + FWHT + sign flips\n"
                "Block size: 128 floats -> 56 bytes"
            )

        return (stats,)


# Node registration dicts
NODE_CLASS_MAPPINGS = {
    "TurboQuantPatch": TurboQuantPatch,
    "TurboQuantInfo": TurboQuantInfo,
}

NODE_DISPLAY_NAME_MAPPINGS = {
    "TurboQuantPatch": "TurboQuant KV Patch",
    "TurboQuantInfo": "TurboQuant Info",
}