kolors got poor results when text encoder quantize[Bug]

[chuck-ma]

Your current environment information

PyTorch version: 2.3.0+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A

OneFlow version: path: ['/root/miniconda3/lib/python3.10/site-packages/oneflow'], version: 0.9.1.dev20240731+cu118, git_commit: f230775, cmake_build_type: Release, rdma: True, mlir: True, enterprise: False Nexfort version: none OneDiff version: 1.2.1.dev10+g72a11f94 OneDiffX version: unknown version

OS: Ubuntu 22.04.1 LTS (x86_64) GCC version: (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0 Clang version: Could not collect CMake version: version 3.29.5 Libc version: glibc-2.35

Python version: 3.10.13 (main, Sep 11 2023, 13:44:35) [GCC 11.2.0] (64-bit runtime) Python platform: Linux-5.15.0-78-generic-x86_64-with-glibc2.35 Is CUDA available: True CUDA runtime version: 11.8.89 CUDA_MODULE_LOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090 Nvidia driver version: 535.104.05 cuDNN version: Probably one of the following: /usr/lib/x86_64-linux-gnu/libcudnn.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.6.0 /usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.6.0 HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 128 On-line CPU(s) list: 0-127 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Platinum 8352V CPU @ 2.10GHz CPU family: 6 Model: 106 Thread(s) per core: 2 Core(s) per socket: 32 Socket(s): 2 Stepping: 6 CPU max MHz: 3500.0000 CPU min MHz: 800.0000 BogoMIPS: 4000.00 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 invpcid_single ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 3 MiB (64 instances) L1i cache: 2 MiB (64 instances) L2 cache: 80 MiB (64 instances) L3 cache: 108 MiB (2 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-31,64-95 NUMA node1 CPU(s): 32-63,96-127 Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

Versions of relevant libraries: [pip3] diffusers==0.30.0.dev0 [pip3] numpy==1.24.1 [pip3] onnxruntime==1.18.1 [pip3] onnxruntime-gpu==1.18.1 [pip3] torch==2.3.0 [pip3] torchvision==0.18.0 [pip3] transformers==4.41.2 [pip3] triton==2.3.0 [conda] numpy 1.24.1 pypi_0 pypi [conda] torch 2.3.0 pypi_0 pypi [conda] torchvision 0.18.0 pypi_0 pypi [conda] triton 2.3.0 pypi_0 pypi

🐛 Describe the bug

import torch
from diffusers import DPMSolverMultistepScheduler, KolorsPipeline

import time
from onediffx import compile_pipe, load_pipe, quantize_pipe, save_pipe
from torch.nn.parameter import Parameter
import bz2
import torch
import base64
import ctypes
from transformers.utils import logging
from typing import List

from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up

class Kernel:
    def __init__(self, code: bytes, function_names: List[str]):
        self.code = code
        self._function_names = function_names
        self._cmodule = LazyKernelCModule(self.code)

        for name in self._function_names:
            setattr(self, name, KernelFunction(self._cmodule, name))

logger = logging.get_logger(__name__)

quantization_code = "$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"

kernels = Kernel(
    bz2.decompress(base64.b64decode(quantization_code)),
    [
        "int4WeightCompression",
        "int4WeightExtractionFloat",
        "int4WeightExtractionHalf",
        "int8WeightExtractionFloat",
        "int8WeightExtractionHalf",
    ],
)

class W8A16Linear(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        inp: torch.Tensor,
        quant_w: torch.Tensor,
        scale_w: torch.Tensor,
        weight_bit_width,
    ):
        ctx.inp_shape = inp.size()
        ctx.weight_bit_width = weight_bit_width
        out_features = quant_w.size(0)
        inp = inp.contiguous().view(-1, inp.size(-1))
        weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
        ctx.weight_shape = weight.size()
        output = inp.mm(weight.t())
        ctx.save_for_backward(inp, quant_w, scale_w)
        return output.view(*(ctx.inp_shape[:-1] + (out_features,)))

    @staticmethod
    def backward(ctx, grad_output: torch.Tensor):
        inp, quant_w, scale_w = ctx.saved_tensors
        weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width)
        grad_output = grad_output.contiguous().view(-1, weight.size(0))
        grad_input = grad_output.mm(weight)
        grad_weight = grad_output.t().mm(inp)
        return (
            grad_input.view(ctx.inp_shape),
            grad_weight.view(ctx.weight_shape),
            None,
            None,
        )

def compress_int4_weight(weight: torch.Tensor):  # (n, m)
    with torch.cuda.device(weight.device):
        n, m = weight.size(0), weight.size(1)
        assert m % 2 == 0
        m = m // 2
        out = torch.empty(n, m, dtype=torch.int8, device="cuda")
        stream = torch.cuda.current_stream()

        gridDim = (n, 1, 1)
        blockDim = (min(round_up(m, 32), 1024), 1, 1)

        kernels.int4WeightCompression(
            gridDim,
            blockDim,
            0,
            stream,
            [
                ctypes.c_void_p(weight.data_ptr()),
                ctypes.c_void_p(out.data_ptr()),
                ctypes.c_int32(n),
                ctypes.c_int32(m),
            ],
        )
        return out

def extract_weight_to_half(
    weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int
):
    assert scale_list.dtype in [torch.half, torch.bfloat16]
    assert weight.dtype in [torch.int8]
    if source_bit_width == 8:
        return weight.to(scale_list.dtype) * scale_list[:, None]
    elif source_bit_width == 4:
        func = (
            kernels.int4WeightExtractionHalf
            if scale_list.dtype == torch.half
            else kernels.int4WeightExtractionBFloat16
        )
    else:
        assert False, "Unsupported bit-width"

    with torch.cuda.device(weight.device):
        n, m = weight.size(0), weight.size(1)
        out = torch.empty(
            n, m * (8 // source_bit_width), dtype=scale_list.dtype, device="cuda"
        )
        stream = torch.cuda.current_stream()

        gridDim = (n, 1, 1)
        blockDim = (min(round_up(m, 32), 1024), 1, 1)

        func(
            gridDim,
            blockDim,
            0,
            stream,
            [
                ctypes.c_void_p(weight.data_ptr()),
                ctypes.c_void_p(scale_list.data_ptr()),
                ctypes.c_void_p(out.data_ptr()),
                ctypes.c_int32(n),
                ctypes.c_int32(m),
            ],
        )
        return out

class QuantizedLinear(torch.nn.Module):
    def __init__(
        self,
        weight_bit_width: int,
        weight,
        bias=None,
        device="cuda",
        dtype=None,
        empty_init=False,
    ):
        super().__init__()
        weight = weight.to(device)  # ensure the weight is on the cuda device
        assert str(weight.device).startswith(
            "cuda"
        ), "The weights that need to be quantified should be on the CUDA device"
        self.weight_bit_width = weight_bit_width
        shape = weight.shape

        if weight is None or empty_init:
            self.weight = torch.empty(
                shape[0],
                shape[1] * weight_bit_width // 8,
                dtype=torch.int8,
                device=device,
            )
            self.weight_scale = torch.empty(shape[0], dtype=dtype, device=device)
        else:
            self.weight_scale = weight.abs().max(dim=-1).values / (
                (2 ** (weight_bit_width - 1)) - 1
            )
            self.weight = torch.round(weight / self.weight_scale[:, None]).to(
                torch.int8
            )
            if weight_bit_width == 4:
                self.weight = compress_int4_weight(self.weight)

        self.weight = Parameter(self.weight.to(device), requires_grad=False)
        self.weight_scale = Parameter(self.weight_scale.to(device), requires_grad=False)
        self.bias = (
            Parameter(bias.to(device), requires_grad=False)
            if bias is not None
            else None
        )

    def forward(self, input):
        output = W8A16Linear.apply(
            input, self.weight, self.weight_scale, self.weight_bit_width
        )
        if self.bias is not None:
            output = output + self.bias
        return output

def quantize_text_encoder(model, weight_bit_width, empty_init=False, device=None):
    """Replace fp16 linear with quantized linear"""
    for layer in model.layers:
        layer.self_attention.query_key_value = QuantizedLinear(
            weight_bit_width=weight_bit_width,
            weight=layer.self_attention.query_key_value.weight,
            bias=layer.self_attention.query_key_value.bias,
            dtype=layer.self_attention.query_key_value.weight.dtype,
            device=(
                layer.self_attention.query_key_value.weight.device
                if device is None
                else device
            ),
            empty_init=empty_init,
        )
        layer.self_attention.dense = QuantizedLinear(
            weight_bit_width=weight_bit_width,
            weight=layer.self_attention.dense.weight,
            bias=layer.self_attention.dense.bias,
            dtype=layer.self_attention.dense.weight.dtype,
            device=(
                layer.self_attention.dense.weight.device if device is None else device
            ),
            empty_init=empty_init,
        )
        layer.mlp.dense_h_to_4h = QuantizedLinear(
            weight_bit_width=weight_bit_width,
            weight=layer.mlp.dense_h_to_4h.weight,
            bias=layer.mlp.dense_h_to_4h.bias,
            dtype=layer.mlp.dense_h_to_4h.weight.dtype,
            device=layer.mlp.dense_h_to_4h.weight.device if device is None else device,
            empty_init=empty_init,
        )
        layer.mlp.dense_4h_to_h = QuantizedLinear(
            weight_bit_width=weight_bit_width,
            weight=layer.mlp.dense_4h_to_h.weight,
            bias=layer.mlp.dense_4h_to_h.bias,
            dtype=layer.mlp.dense_4h_to_h.weight.dtype,
            device=layer.mlp.dense_4h_to_h.weight.device if device is None else device,
            empty_init=empty_init,
        )

    return model

device = torch.device("cuda")
model = "Kwai-Kolors/Kolors-diffusers"
oneflow_load = True
oneflow_save = True

nexfort_compiler_config = None
nexfort_quantize_config = None

class KolorsGenerator:
    def __init__(
        self,
        model,
        compiler_config=None,
        quantize_config=None,
        compiler="none",
        enable_compiler=True,
    ):
        pipe = KolorsPipeline.from_pretrained(
            model, torch_dtype=torch.float16, variant="fp16"
        ).to(device)

        start_time = time.time()
        pipe.text_encoder = pipe.text_encoder.to("cuda")

        quantize_text_encoder(pipe.text_encoder.encoder, 4)
        torch.cuda.empty_cache()
        end_time = time.time()

        inference_time = end_time - start_time
        logger.info(f"text encoder quantize time: {inference_time:.4f} seconds")

        pipe.unet = pipe.unet.to("cuda")
        pipe.vae = pipe.vae.to("cuda")
        pipe.scheduler = DPMSolverMultistepScheduler.from_config(
            pipe.scheduler.config, use_karras_sigmas=True
        )

        self.pipe = pipe

        self.enable_compiler = enable_compiler

        if self.enable_compiler:
            if compiler == "nexfort":
                if compiler_config:
                    print("nexfort backend compile...")
                    self.pipe = self.compile_pipe(self.pipe, compiler_config)

                if quantize_config:
                    print("nexfort backend quant...")
                    self.pipe = self.quantize_pipe(self.pipe, quantize_config)
            elif compiler == "oneflow":
                logger.info("oneflow backend compile...")
                self.pipe = compile_pipe(self.pipe, ignores=["text_encoder", "vae"])

            self.warmup()

    def warmup(
        self,
        gen_args={
            "prompt": '一张瓢虫的照片，微距，变焦，高质量，电影，拿着一个牌子，写着"可图"',
            "num_inference_steps": 25,
            "height": 1024,
            "width": 1024,
            "guidance_scale": 6.5,
        },
        warmup_iterations=1,
    ):
        warmup_args = gen_args.copy()

        warmup_args["generator"] = torch.Generator(device=device).manual_seed(0)

        logger.info("Starting warmup...")
        start_time = time.time()
        if oneflow_load:
            load_pipe(self.pipe, dir="cached_pipe")

        for _ in range(warmup_iterations):
            self.pipe(**warmup_args)

        if oneflow_save:
            save_pipe(self.pipe, dir="cached_pipe")
        end_time = time.time()
        logger.info(f"Warmup time: {end_time - start_time:.2f} seconds")

    def generate(self, gen_args):
        # gen_args["generator"] = torch.Generator(device=device).manual_seed(0)

        # Run the model
        start_time = time.time()
        images = self.pipe(**gen_args).images
        end_time = time.time()
        inference_time = end_time - start_time
        logger.info(f"paint_pipe inference time: {inference_time:.2f} seconds")
        torch.cuda.empty_cache()

        return images

    def compile_pipe(self, pipe, compiler_config):
        options = compiler_config
        pipe = compile_pipe(
            pipe, backend="nexfort", options=options, fuse_qkv_projections=True
        )
        return pipe

    def quantize_pipe(self, pipe, quantize_config):
        pipe = quantize_pipe(pipe, ignores=[], **quantize_config)
        return pipe

kolors = KolorsGenerator(
    model,
    nexfort_compiler_config,
    nexfort_quantize_config,
    compiler="oneflow",
    enable_compiler=True,
)

gen_args = {
        "prompt": '一个女孩',
        "num_inference_steps": 25,
        "height": 576,
        "width": 1024,
        "guidance_scale": 6.5,
        "num_images_per_prompt":4,
    }

res_images = paint_pipe.generate(gen_args=gen_args)

when compiler is disabled, result looks good:

However, when enable compiler, result looks quite bad:

[lixiang007666]

Try updating to the new version of nextort. @chuck-ma

[chuck-ma]

@lixiang007666 I will try it. But I use oneflow, because it's faster.

[lixiang007666]

@lixiang007666 I will try it. But I use oneflow, because it's faster.

For the unet architecture series of models, oneflow is currently faster and easier to use.

Nexfort is currently more suitable for DiT. Nexfort backend will continue to be optimized in the future.

@chuck-ma

[chuck-ma]

Anyway to speed up warm up time when using nextort? It seems that it uses only a cpu kernel

[chuck-ma]

It gets errors...

AttributeError: 'ChatGLMModel' object has no attribute '_deployable_module_dpl_graph'

My nexfort version: nexfort 0.1.dev260

[lixiang007666]

It gets errors...

AttributeError: 'ChatGLMModel' object has no attribute '_deployable_module_dpl_graph'

My nexfort version:

nexfort 0.1.dev260

Thanks for finding this issue, I need to try to reproduce it.

[lixiang007666]

@chuck-ma

        start_time = time.time()
        pipe.text_encoder = pipe.text_encoder.to("cuda")

        quantize_text_encoder(pipe.text_encoder.encoder, 4)
        torch.cuda.empty_cache()
        end_time = time.time()

        inference_time = end_time - start_time
        logger.info(f"text encoder quantize time: {inference_time:.4f} seconds")

I suggest that you remove the quantize_text_encoder() lines of code.

If there are still issues, I will continue to follow up.

[lixiang007666]

Anyway to speed up warm up time when using nextort? It seems that it uses only a cpu kernel

@chuck-ma And, for the nexfort backend, the issue of prolonged tuning time after the first compilation can be resolved by using cache.

https://github.com/siliconflow/onediff/tree/main/src/onediff/infer_compiler/backends/nexfort#compilation-cache-speeds-up-recompilation

Warmup time after use cache:

env:

nexfort==0.1.dev261
torch==2.4.0

[chuck-ma]

@chuck-ma

        start_time = time.time()
        pipe.text_encoder = pipe.text_encoder.to("cuda")

        quantize_text_encoder(pipe.text_encoder.encoder, 4)
        torch.cuda.empty_cache()
        end_time = time.time()

        inference_time = end_time - start_time
        logger.info(f"text encoder quantize time: {inference_time:.4f} seconds")

I suggest that you remove the quantize_text_encoder() lines of code.

If there are still issues, I will continue to follow up.

It seems that if i quantize text encoder, it will also speed up generating images, which also saves my cuda memory. I will be appreciate if onediff could support speed up even text encoder is quantized