Merge remote-tracking branch 'origin/main' into merge

bitsandbytes/research/__init__.py

+from . import nn
+from .autograd._functions import (
+    switchback_bnb,
+    matmul_fp8_global,
+    matmul_fp8_mixed,
+)

bitsandbytes/research/autograd/_functions.py

+import operator
+import warnings
+from dataclasses import dataclass
+from functools import reduce  # Required in Python 3
+
+import torch
+
+import bitsandbytes.functional as F
+
+from bitsandbytes.autograd._functions import MatmulLtState, GlobalOutlierPooler
+
+
+# math.prod not compatible with python < 3.8
+def prod(iterable):
+    return reduce(operator.mul, iterable, 1)
+
+tensor = torch.Tensor
+
+class MatMulFP8Mixed(torch.autograd.Function):
+    # forward is the same, but we added the fallback for pre-turing GPUs
+    # backward is mostly the same, but adds one extra clause (see "elif state.CxB is not None")
+
+    @staticmethod
+    def forward(ctx, A, B, out=None, fw_code=None, bw_code=None, bsz=1024, bsz2=1024):
+        # default of pytorch behavior if inputs are empty
+        ctx.is_empty = False
+        if prod(A.shape) == 0:
+            ctx.is_empty = True
+            ctx.A = A
+            ctx.B = B
+
+            B_shape = B.shape
+            if A.shape[-1] == B_shape[0]:
+                return torch.empty(A.shape[:-1] + B_shape[1:], dtype=A.dtype, device=A.device)
+            else:
+                return torch.empty(A.shape[:-1] + B_shape[:1], dtype=A.dtype, device=A.device)
+
+        # 1. Dequantize
+        # 2. MatmulnN
+        cA, state = F.quantize_blockwise(A, code=fw_code, blocksize=bsz)
+        fp8A = F.dequantize_blockwise(cA, state, blocksize=bsz).to(A.dtype)
+
+        cB, state = F.quantize(B.float(), code=fw_code)
+        fp8B = F.dequantize(cB, state).to(B.dtype)
+
+        output = torch.matmul(fp8A, fp8B)
+
+        # output is half
+
+        # 3. Save state
+        ctx.fw_code = fw_code
+        ctx.bw_code = bw_code
+        ctx.bsz = bsz
+        ctx.bsz2 = bsz2
+        ctx.dtype_A, ctx.dtype_B = A.dtype, B.dtype
+
+        if any(ctx.needs_input_grad[:2]):
+            # NOTE: we send back A, and re-quant.
+            ctx.tensors = (A, fp8B)
+        else:
+            ctx.tensors = (None, None)
+
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.is_empty:
+            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, None, None, None, None
+
+        req_gradA, req_gradB, _, _, _, _, _ = ctx.needs_input_grad
+        A, B = ctx.tensors
+
+        grad_A, grad_B = None, None
+
+        # TODO: Fix blocksize to be output_dim
+        cgrad_out, state = F.quantize_blockwise(grad_output, code=ctx.bw_code, blocksize=ctx.bsz2)
+        fp8out = F.dequantize_blockwise(cgrad_out, state, blocksize=ctx.bsz2).to(grad_output.dtype)
+
+        # cgrad_output_2, state_2 = F.quantize(grad_output.float(), code=ctx.bw_code)
+        # fp8out_2 = F.dequantize(cgrad_output_2, state_2).to(grad_output.dtype)
+
+        # grad_output_reshape = grad_output.reshape(-1, grad_output.shape[-1]).contiguous()
+        # fp8grad_transpose, stategrad_transpose = F.vectorwise_quant(grad_output_reshape, dim=0, quant_type='vector')
+        # fp8out_transpose = (fp8grad_transpose / 7) * stategrad_transpose
+        # fp8out_transpose = fp8out_transpose.view(grad_output.shape[0], grad_output.shape[1], grad_output.shape[2])
+
+        # not supported by PyTorch. TODO: create work-around
+        if req_gradA: 
+            grad_A = torch.matmul(fp8out, B.t().to(fp8out.dtype)).to(A.dtype)
+
+        if req_gradB:
+            if len(A.shape) == 3:
+                At = A.transpose(2, 1).contiguous()
+            else:
+                At = A.transpose(1, 0).contiguous()
+            # cA, state = F.quantize(At.float(), code=ctx.fw_code)
+            # fp8At = F.dequantize(cA, state).to(A.dtype)
+            grad_B = torch.matmul(At.to(grad_output.dtype), grad_output).to(B.dtype)
+
+        return grad_A, grad_B, None, None, None, None, None
+
+
+class MatMulFP8Global(torch.autograd.Function):
+    # forward is the same, but we added the fallback for pre-turing GPUs
+    # backward is mostly the same, but adds one extra clause (see "elif state.CxB is not None")
+
+    @staticmethod
+    def forward(ctx, A, B, out=None, fw_code=None, bw_code=None, bsz=1024, bsz2=1024):
+        # default of pytorch behavior if inputs are empty
+        ctx.is_empty = False
+        if prod(A.shape) == 0:
+            ctx.is_empty = True
+            ctx.A = A
+            ctx.B = B
+
+            B_shape = B.shape
+            if A.shape[-1] == B_shape[0]:
+                return torch.empty(A.shape[:-1] + B_shape[1:], dtype=A.dtype, device=A.device)
+            else:
+                return torch.empty(A.shape[:-1] + B_shape[:1], dtype=A.dtype, device=A.device)
+
+        # 1. Dequantize
+        # 2. MatmulnN
+        cA, state = F.quantize(A.float(), code=fw_code)
+        fp8A = F.dequantize(cA, state).to(A.dtype)
+
+        cB, state = F.quantize(B.float(), code=fw_code)
+        fp8B = F.dequantize(cB, state).to(B.dtype)
+
+        output = torch.matmul(fp8A, fp8B)
+
+        # output is half
+
+        # 3. Save state
+        ctx.fw_code = fw_code
+        ctx.bw_code = bw_code
+        ctx.bsz = bsz
+        ctx.bsz2 = bsz2
+        ctx.dtype_A, ctx.dtype_B = A.dtype, B.dtype
+
+        if any(ctx.needs_input_grad[:2]):
+            # NOTE: we send back A, and re-quant.
+            ctx.tensors = (A, fp8B)
+        else:
+            ctx.tensors = (None, None)
+
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.is_empty:
+            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, None, None, None, None
+
+        req_gradA, req_gradB, _, _, _, _, _ = ctx.needs_input_grad
+        A, B = ctx.tensors
+
+        grad_A, grad_B = None, None
+
+        # TODO: Fix blocksize to be output_dim
+        cgrad_out, state = F.quantize(grad_output.float(), code=ctx.bw_code)
+        fp8out = F.dequantize(cgrad_out, state).to(grad_output.dtype)
+
+        # cgrad_output_2, state_2 = F.quantize(grad_output.float(), code=ctx.bw_code)
+        # fp8out_2 = F.dequantize(cgrad_output_2, state_2).to(grad_output.dtype)
+
+        # grad_output_reshape = grad_output.reshape(-1, grad_output.shape[-1]).contiguous()
+        # fp8grad_transpose, stategrad_transpose = F.vectorwise_quant(grad_output_reshape, dim=0, quant_type='vector')
+        # fp8out_transpose = (fp8grad_transpose / 7) * stategrad_transpose
+        # fp8out_transpose = fp8out_transpose.view(grad_output.shape[0], grad_output.shape[1], grad_output.shape[2])
+
+        # not supported by PyTorch. TODO: create work-around
+        if req_gradA: 
+            grad_A = torch.matmul(fp8out, B.t().to(fp8out.dtype)).to(A.dtype)
+
+        if req_gradB:
+            if len(A.shape) == 3:
+                At = A.transpose(2, 1).contiguous()
+            else:
+                At = A.transpose(1, 0).contiguous()
+            cA, state = F.quantize(At.float(), code=ctx.fw_code)
+            fp8At = F.dequantize(cA, state).to(A.dtype)
+            grad_B = torch.matmul(fp8At.to(fp8out.dtype), fp8out).to(B.dtype)
+
+        return grad_A, grad_B, None, None, None, None, None
+
+
+class SwitchBackBnb(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, A, B, out=None, bias=None, state=MatmulLtState()):
+        # default to pytorch behavior if inputs are empty
+        ctx.is_empty = False
+        if prod(A.shape) == 0:
+            ctx.is_empty = True
+            ctx.A = A
+            ctx.B = B
+            ctx.bias = bias
+            if A.shape[-1] == B.shape[0]:
+                return torch.empty(A.shape[:-1]+B.shape[1:], dtype=A.dtype, device=A.device)
+            else:
+                return torch.empty(A.shape[:-1]+B.shape[:1], dtype=A.dtype, device=A.device)
+
+        # 1. Quantize A
+        # 2. Quantize B
+        # 3. Matmul
+        # 4. Mixed-precision decomposition matmul
+        # 5. Save state
+        formatB = state.formatB
+        input_shape = A.shape
+        if state.outlier_pool is None:
+            state.outlier_pool = GlobalOutlierPooler.get_instance()
+
+        # Cast A to fp16
+        if A.dtype != torch.float16:
+            warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization")
+
+        # 1. Quantize A
+        if len(A.shape) == 3:
+            A = A.view(-1, A.shape[-1]).contiguous()
+        CA, CAt, SCA, SCAt, coo_tensorA = F.double_quant(
+            A.to(torch.float16), threshold=state.threshold
+        )
+
+        if state.threshold > 0.0 and coo_tensorA is not None:
+            if state.has_fp16_weights:
+                idx = torch.unique(coo_tensorA.colidx).long()
+                CA[:, idx] = 0
+                CAt[:, idx] = 0
+                subA = A[:, idx]
+                state.subB = B[:, idx].t().contiguous()
+                state.idx = idx
+            else:
+                if state.CxB is None:
+                    # B in in 8-bit row-major, we can transform it back to 16-bit to extract outlier dimensions
+                    # we also need to convert it to the turing/ampere format
+                    state.CxB, state.SB = F.transform(state.CB, to_order=formatB)
+        else:
+            #print('A shape', A.shape)
+            if not state.has_fp16_weights and state.CxB is None:
+                state.CxB, state.SB = F.transform(state.CB, to_order=formatB)
+            subA = None
+
+        # 2. Quantize B
+        if state.has_fp16_weights:
+            #print('B shape', B.shape)
+            has_grad = True if (getattr(B, "grad", None) is not None) else False
+            is_transposed = not B.is_contiguous() and B.shape[0] == B.stride(1)
+            if is_transposed:
+                B = B.contiguous()
+
+            if (state.is_training and not has_grad) or state.CxB is None:
+                state.reset_grads()
+                (
+                    CB,
+                    state.CBt,
+                    state.SCB,
+                    state.SCBt,
+                    coo_tensorB,
+                ) = F.double_quant(B.to(torch.float16))
+                state.CxB, state.SB = F.transform(CB, to_order=formatB)
+        else:
+            has_grad = False
+
+        if coo_tensorA is not None and not state.has_fp16_weights:
+            # extract outliers
+
+            outlier_idx = torch.unique(coo_tensorA.colidx)
+            state.idx = outlier_idx
+            # state.outlier_pool.add_outliers(outlier_idx, A.shape[-1])
+            # if state.use_pool and state.outlier_pool.model_dim == A.shape[-1]:
+            #    # do not use pool for 2nd FFN layer
+            #    state.idx = state.outlier_pool.get_current_outlier_idx().to(A.device)
+            # else:
+            #    state.idx = outlier_idx
+            outliers = F.extract_outliers(state.CxB, state.SB, state.idx.int())
+            state.subB = (
+                (outliers * state.SCB.view(-1, 1) / 127.0)
+                .t()
+                .contiguous()
+                .to(A.dtype)
+            )
+            CA[:, state.idx.long()] = 0
+            CAt[:, state.idx.long()] = 0
+            subA = A[:, state.idx.long()]
+
+        shapeB = state.SB[0]
+
+        if len(input_shape) == 3:
+            output_shape = (input_shape[0], input_shape[1], shapeB[0])
+        else:
+            output_shape = (input_shape[0], shapeB[0])
+
+        # 3. Matmul
+        C32A, SA = F.transform(CA, "col32")
+        out32, Sout32 = F.igemmlt(C32A, state.CxB, SA, state.SB)
+        # we apply the fused bias here
+
+        if bias is None or bias.dtype == torch.float16:
+            output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=bias)
+            output = output.to(A.dtype)
+        else:  # apply bias separately
+            output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=None)
+            output = output.to(A.dtype).add_(bias)
+
+        # 4. Mixed-precision decomposition matmul
+        if coo_tensorA is not None and subA is not None:
+            output += torch.matmul(subA, state.subB)
+
+        # 5. Save state
+        ctx.state = state
+
+        ctx.formatB = formatB
+        ctx.grad_shape = input_shape
+        ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype
+
+        if any(ctx.needs_input_grad[:2]):
+            ctx.tensors = (CAt, subA, A)
+            ctx.tensor_states = (SCAt, state.idx)
+        else:
+            ctx.tensors = [None, None, None]
+            ctx.tensor_states = (None, None)
+            ctx.save_for_backward(None, None)
+
+
+        clone_func = torch.clone if len(output_shape) == 3 else lambda x : x
+        return clone_func(output.view(output_shape))
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.is_empty:
+            bias_grad = (None if ctx.bias is None else torch.zeros_like(ctx.bias))
+            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
+        req_gradA, req_gradB, _, req_gradBias, _ = ctx.needs_input_grad
+        CAt, subA, A = ctx.tensors
+        SCAt, idx = ctx.tensor_states
+        formatB = ctx.formatB
+        state = ctx.state
+        grad_A = grad_B = grad_bias = None
+
+        if req_gradBias:
+            # compute grad_bias first before changing grad_output dtype
+            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)
+
+        # Cast grad_output to fp16
+        if len(grad_output.shape) == 3:
+            grad_output = grad_output.reshape(
+                -1, grad_output.shape[-1]
+            ).contiguous()
+
+        Cgrad, Cgradt, SCgrad, SCgradt, coo_tensor = F.double_quant(grad_output.to(torch.float16))
+
+        if req_gradB:
+            # print('back A shape', A.shape)
+            # print('grad output t shape', grad_output.t().shape)
+            grad_B = torch.matmul(grad_output.t(), A)
+
+        if req_gradA:
+            if state.CBt is not None:
+                C32grad, Sgrad = F.transform(Cgrad, "col32")
+                if state.CxBt is None:
+                    state.CxBt, state.SBt = F.transform(
+                        state.CBt, to_order=formatB, transpose=True
+                    )
+                # print('back B shape', state.CxBt.shape)
+                # print('back grad shape', C32grad.shape)
+                gradA32, SgradA32 = F.igemmlt(C32grad, state.CxBt, Sgrad, state.SBt)
+                grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape).to(ctx.dtype_A)
+
+            elif state.CB is not None:
+                CB = state.CB.to(ctx.dtype_A, copy=True).mul_(state.SCB.unsqueeze(1).mul(1. / 127.0))
+                grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A)
+            else:
+                raise Exception('State must contain either CBt or CB matrix for backward')
+
+        return grad_A, grad_B, None, grad_bias, None
+
+def get_block_sizes(input_matrix, weight_matrix):
+    input_features = input_matrix.shape[-1]
+    output_features = (weight_matrix.shape[0] if weight_matrix.shape[1] == input_features else weight_matrix.shape[1])
+    array = [4096, 2048, 1024, 512, 256, 128, 64, 0]
+    bsz, bsz2 = 1024, 1024
+    for i, k in enumerate(array):
+        if input_features > array[i + 1]:
+            bsz = k
+            break
+    for i, k in enumerate(array):
+        if output_features > array[i + 1]:
+            bsz2 = k
+            break
+
+    return bsz, bsz2
+
+def matmul_fp8_global(A: tensor, B: tensor, fw_code: tensor, bw_code: tensor, out: tensor = None, bsz : int = -1, bsz2 : int = -1):
+    if bsz == -1 or bsz2 == -1: bsz, bsz2 = get_block_sizes(A, B)
+    return MatMulFP8Global.apply(A, B, out, fw_code, bw_code, bsz, bsz2)
+
+def matmul_fp8_mixed(A: tensor, B: tensor, fw_code: tensor, bw_code: tensor, out: tensor = None, bsz : int = -1, bsz2 : int = -1):
+    if bsz == -1 or bsz2 == -1: bsz, bsz2 = get_block_sizes(A, B)
+    return MatMulFP8Mixed.apply(A, B, out, fw_code, bw_code, bsz, bsz2)
+
+def switchback_bnb(
+    A: tensor,
+    B: tensor,
+    out: tensor = None,
+    state: MatmulLtState = None,
+    threshold=0.0,
+    bias=None
+):
+    state = state or MatmulLtState()
+    if threshold > 0.0:
+        state.threshold = threshold
+    return SwitchBackBnb.apply(A, B, out, bias, state)

bitsandbytes/research/nn/__init__.py

+from .modules import LinearFP8Mixed, LinearFP8Global

bitsandbytes/research/nn/modules.py

+from typing import Optional, TypeVar, Union, overload
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, device, dtype, nn
+
+import bitsandbytes as bnb
+from bitsandbytes.optim import GlobalOptimManager
+from bitsandbytes.utils import OutlierTracer, find_outlier_dims
+
+T = TypeVar("T", bound="torch.nn.Module")
+
+
+class LinearFP8Mixed(nn.Linear):
+    def __init__(self, input_features, output_features, bias=True):
+        super().__init__(input_features, output_features, bias)
+        self.bw_code = None
+        self.fw_code = None
+        array = [4096, 2048, 1024, 512, 256, 128, 64, 0]
+        for i, k in enumerate(array):
+            if input_features > array[i + 1]:
+                self.bsz = k
+                break
+        for i, k in enumerate(array):
+            if output_features > array[i + 1]:
+                self.bsz2 = k
+                break
+
+    def forward(self, x: torch.Tensor):
+        if self.fw_code is None:
+            self.bw_code = bnb.functional.create_fp8_map(True, 5, 2, 8).to(x.device)
+            self.fw_code = bnb.functional.create_fp8_map(True, 4, 3, 8).to(x.device)
+
+        out = bnb.research.matmul_fp8_mixed(x, self.weight.t(), fw_code=self.fw_code, bw_code=self.bw_code, bsz=self.bsz, bsz2=self.bsz2)
+        if self.bias is not None:
+            out += self.bias
+
+        return out
+
+class LinearFP8Global(nn.Linear):
+    def __init__(self, input_features, output_features, bias=True):
+        super().__init__(input_features, output_features, bias)
+        self.bw_code = None
+        self.fw_code = None
+        array = [4096, 2048, 1024, 512, 256, 128, 64, 0]
+        for i, k in enumerate(array):
+            if input_features > array[i + 1]:
+                self.bsz = k
+                break
+        for i, k in enumerate(array):
+            if output_features > array[i + 1]:
+                self.bsz2 = k
+                break
+
+    def forward(self, x: torch.Tensor):
+        if self.fw_code is None:
+            self.bw_code = bnb.functional.create_fp8_map(True, 5, 2, 8).to(x.device)
+            self.fw_code = bnb.functional.create_fp8_map(True, 4, 3, 8).to(x.device)
+
+        out = bnb.matmul_fp8_global(x, self.weight.t(), fw_code=self.fw_code, bw_code=self.bw_code, bsz=self.bsz, bsz2=self.bsz2)
+        if self.bias is not None:
+            out += self.bias
+
+        return out

bitsandbytes/triton/dequantize_rowwise.py

+import math
+import torch
+import time
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def dequantize_rowwise(x: torch.Tensor, state_x: torch.Tensor): return None
+else:
+
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+    # rowwise quantize
+
+    # TODO: autotune this better.
+    @triton.autotune(
+            configs=[
+                triton.Config({}, num_stages=1, num_warps=8),
+                triton.Config({}, num_stages=2, num_warps=8),
+                triton.Config({}, num_stages=4, num_warps=8),
+                triton.Config({}, num_stages=8, num_warps=8),
+                triton.Config({}, num_stages=1),
+                triton.Config({}, num_stages=2),
+                triton.Config({}, num_stages=4),
+                triton.Config({}, num_stages=8),
+                triton.Config({}, num_warps=1),
+                triton.Config({}, num_warps=2),
+                triton.Config({}, num_warps=4),
+                triton.Config({}, num_warps=8),
+            ],
+            key=['n_elements']
+    )
+    @triton.jit
+    def _dequantize_rowwise(
+        x_ptr,
+        state_x,
+        output_ptr,
+        inv_127,
+        n_elements,
+        BLOCK_SIZE: tl.constexpr,
+        P2: tl.constexpr,
+    ):
+        pid = tl.program_id(axis=0)
+        block_start = pid * BLOCK_SIZE
+        arange = tl.arange(0, P2)
+        offsets = block_start + arange
+        row_mask = arange < BLOCK_SIZE
+        x = tl.load(x_ptr + offsets, mask=row_mask)
+        max_val = tl.load(state_x + pid)
+        output = max_val * x * inv_127
+        tl.store(output_ptr + offsets, output, mask=row_mask)
+        
+
+    def dequantize_rowwise(x: torch.Tensor, state_x: torch.Tensor):
+        output = torch.empty(*x.shape, device=x.device, dtype=torch.float16)
+
+        P2 = int(2 ** (math.ceil(math.log2(x.shape[1]))))
+
+        assert x.is_cuda and output.is_cuda
+        n_elements = output.numel()
+        grid = lambda meta: (x.shape[0],)
+        _dequantize_rowwise[grid](x, state_x, output, 1./127, n_elements, BLOCK_SIZE=x.shape[1], P2=P2)
+        return output

bitsandbytes/triton/int8_matmul_mixed_dequanitze.py

+import torch
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def int8_matmul_mixed_dequanitze(a, b, state_x, state_w, bias): return None
+else:
+
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+
+    # This is a matmul kernel based on triton.ops.matmul
+    # It is modified to support rowwise quantized input and global quantized weight
+    # It's purpose is fused matmul then dequantize
+    # It does support bias.
+
+    def init_to_zero(name):
+        return lambda nargs: nargs[name].zero_()
+
+    def get_configs_io_bound():
+        configs = []
+        for num_stages in [2, 3, 4, 5, 6]:
+            for block_m in [16, 32]:
+                for block_k in [32, 64]:
+                    for block_n in [32, 64, 128, 256]:
+                        num_warps = 2 if block_n <= 64 else 4
+                        configs.append(
+                            triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': 1},
+                                          num_stages=num_stages, num_warps=num_warps))
+                        # split_k
+                        for split_k in [2, 4, 8, 16]:
+                            configs.append(triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': split_k},
+                                                         num_stages=num_stages, num_warps=num_warps, pre_hook=init_to_zero('C')))
+        return configs
+
+
+    @triton.autotune(
+        configs=[
+            # basic configs for compute-bound matmuls
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=5, num_warps=2),
+            # good for int8
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=5, num_warps=2),
+        ] + get_configs_io_bound(),
+        key=['M', 'N', 'K'],
+        prune_configs_by={
+            'early_config_prune': early_config_prune,
+            'perf_model': estimate_matmul_time,
+            'top_k': 10
+        },
+    )
+    @triton.heuristics({
+        'EVEN_K': lambda args: args['K'] % (args['BLOCK_K'] * args['SPLIT_K']) == 0,
+    })
+    @triton.jit
+    def _int8_matmul_mixed_dequantize(A, B, C, bias, state_x_ptr, state_w_ptr, M, N, K, divfactor: tl.constexpr, has_bias : tl.constexpr,
+                stride_am, stride_ak,
+                stride_bk, stride_bn,
+                stride_cm, stride_cn,
+                BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr,
+                GROUP_M: tl.constexpr, SPLIT_K: tl.constexpr, EVEN_K: tl.constexpr,
+                ACC_TYPE: tl.constexpr
+                ):
+        # matrix multiplication
+        pid = tl.program_id(0)
+        pid_z = tl.program_id(1)
+        grid_m = tl.cdiv(M, BLOCK_M)
+        grid_n = tl.cdiv(N, BLOCK_N)
+        # re-order program ID for better L2 performance
+        width = GROUP_M * grid_n
+        group_id = pid // width
+        group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
+        pid_m = group_id * GROUP_M + (pid % group_size)
+        pid_n = (pid % width) // (group_size)
+        # do matrix multiplication
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+        ram = tl.max_contiguous(tl.multiple_of(rm % M, BLOCK_M), BLOCK_M)
+        rbn = tl.max_contiguous(tl.multiple_of(rn % N, BLOCK_N), BLOCK_N)
+        rk = pid_z * BLOCK_K + tl.arange(0, BLOCK_K)
+        # pointers
+        A = A + (ram[:, None] * stride_am + rk[None, :] * stride_ak)
+        B = B + (rk[:, None] * stride_bk + rbn[None, :] * stride_bn)
+
+        # rematerialize rm and rn to save registers
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+
+        w_factor = tl.load(state_w_ptr)
+        x_factor = tl.load(state_x_ptr + ram)[:, None]
+
+        # acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=ACC_TYPE)
+        acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.int32)
+        for k in range(0, tl.cdiv(K, BLOCK_K * SPLIT_K)):
+            if EVEN_K:
+                a = tl.load(A)
+                b = tl.load(B)
+            else:
+                k_remaining = K - k * (BLOCK_K * SPLIT_K)
+                a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.)
+                b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.)
+            acc += tl.dot(a, b)
+            A += BLOCK_K * SPLIT_K * stride_ak
+            B += BLOCK_K * SPLIT_K * stride_bk
+        
+        acc = (w_factor * (x_factor * (acc * divfactor)))
+        acc = acc.to(C.dtype.element_ty)
+
+        # conditionally add bias
+        if has_bias:
+            bias = tl.load(bias + rn).to(C.dtype.element_ty)
+            acc = acc + bias[None, :]
+
+        C = C + (rm[:, None] * stride_cm + rn[None, :] * stride_cn)
+        mask = (rm < M)[:, None] & (rn < N)[None, :]
+        # handles write-back with reduction-splitting
+        if SPLIT_K == 1:
+            tl.store(C, acc, mask=mask)
+        else:
+            tl.atomic_add(C, acc, mask=mask)
+
+
+    def int8_matmul_mixed_dequanitze(a, b, state_x, state_w, bias):
+        device = a.device
+        divfactor = 1. / (127. * 127.)
+        has_bias = 0 if bias is None else 1
+        # handle non-contiguous inputs if necessary
+        if a.stride(0) > 1 and a.stride(1) > 1:
+            a = a.contiguous()
+        if b.stride(0) > 1 and b.stride(1) > 1:
+            b = b.contiguous()
+        # checks constraints
+        assert a.shape[1] == b.shape[0], "incompatible dimensions"
+        M, K = a.shape
+        _, N = b.shape
+        # allocates output
+        c = torch.empty((M, N), device=device, dtype=torch.float16)
+        # accumulator types
+        ACC_TYPE = tl.float32 #if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32
+        # launch int8_matmul_mixed_dequantize kernel
+        grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']), META['SPLIT_K'])
+        _int8_matmul_mixed_dequantize[grid](a, b, c, bias, state_x, state_w, M, N, K, divfactor, has_bias,
+                        a.stride(0), a.stride(1),
+                        b.stride(0), b.stride(1),
+                        c.stride(0), c.stride(1),
+                        GROUP_M=8, ACC_TYPE=ACC_TYPE)
+        return c

bitsandbytes/triton/int8_matmul_rowwise_dequantize.py

+import torch
+
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias): return None
+else:
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+    # This is a matmul kernel based on triton.ops.matmul
+    # It is modified to support rowwise quantized input and columnwise quantized weight
+    # It's purpose is fused matmul then dequantize
+    # It does support bias.
+
+    def init_to_zero(name):
+        return lambda nargs: nargs[name].zero_()
+
+
+    def get_configs_io_bound():
+        configs = []
+        for num_stages in [2, 3, 4, 5, 6]:
+            for block_m in [16, 32]:
+                for block_k in [32, 64]:
+                    for block_n in [32, 64, 128, 256]:
+                        num_warps = 2 if block_n <= 64 else 4
+                        configs.append(
+                            triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': 1},
+                                          num_stages=num_stages, num_warps=num_warps))
+                        # split_k
+                        for split_k in [2, 4, 8, 16]:
+                            configs.append(triton.Config({'BLOCK_M': block_m, 'BLOCK_N': block_n, 'BLOCK_K': block_k, 'SPLIT_K': split_k},
+                                                         num_stages=num_stages, num_warps=num_warps, pre_hook=init_to_zero('C')))
+        return configs
+
+
+    @triton.autotune(
+        configs=[
+            # basic configs for compute-bound matmuls
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 32, 'SPLIT_K': 1}, num_stages=5, num_warps=2),
+            # good for int8
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=3, num_warps=8),
+            triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'BLOCK_K': 128, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 128, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=4, num_warps=4),
+            triton.Config({'BLOCK_M': 64, 'BLOCK_N': 32, 'BLOCK_K': 64, 'SPLIT_K': 1}, num_stages=5, num_warps=2),
+        ] + get_configs_io_bound(),
+        key=['M', 'N', 'K'],
+        prune_configs_by={
+            'early_config_prune': early_config_prune,
+            'perf_model': estimate_matmul_time,
+            'top_k': 10
+        },
+    )
+    @triton.heuristics({
+        'EVEN_K': lambda args: args['K'] % (args['BLOCK_K'] * args['SPLIT_K']) == 0,
+    })
+    @triton.jit
+    def _int8_matmul_rowwise_dequantize(A, B, C, bias, state_x_ptr, state_w_ptr, M, N, K, divfactor, has_bias : tl.constexpr,
+                stride_am, stride_ak,
+                stride_bk, stride_bn,
+                stride_cm, stride_cn,
+                BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr,
+                GROUP_M: tl.constexpr, SPLIT_K: tl.constexpr, EVEN_K: tl.constexpr,
+                ACC_TYPE: tl.constexpr
+                ):
+        # matrix multiplication
+        pid = tl.program_id(0)
+        pid_z = tl.program_id(1)
+        grid_m = tl.cdiv(M, BLOCK_M)
+        grid_n = tl.cdiv(N, BLOCK_N)
+        # re-order program ID for better L2 performance
+        width = GROUP_M * grid_n
+        group_id = pid // width
+        group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
+        pid_m = group_id * GROUP_M + (pid % group_size)
+        pid_n = (pid % width) // (group_size)
+        # do matrix multiplication
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+        ram = tl.max_contiguous(tl.multiple_of(rm % M, BLOCK_M), BLOCK_M)
+        rbn = tl.max_contiguous(tl.multiple_of(rn % N, BLOCK_N), BLOCK_N)
+        rk = pid_z * BLOCK_K + tl.arange(0, BLOCK_K)
+        # pointers
+        A = A + (ram[:, None] * stride_am + rk[None, :] * stride_ak)
+        B = B + (rk[:, None] * stride_bk + rbn[None, :] * stride_bn)
+
+        # rematerialize rm and rn to save registers
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+
+        w_factor = tl.load(state_w_ptr + rbn)[None, :]
+        x_factor = tl.load(state_x_ptr + ram)[:, None]
+
+        # acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=ACC_TYPE)
+        acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.int32)
+        for k in range(0, tl.cdiv(K, BLOCK_K * SPLIT_K)):
+            if EVEN_K:
+                a = tl.load(A)
+                b = tl.load(B)
+            else:
+                k_remaining = K - k * (BLOCK_K * SPLIT_K)
+                a = tl.load(A, mask=rk[None, :] < k_remaining, other=0.)
+                b = tl.load(B, mask=rk[:, None] < k_remaining, other=0.)
+            acc += tl.dot(a, b)
+            A += BLOCK_K * SPLIT_K * stride_ak
+            B += BLOCK_K * SPLIT_K * stride_bk
+        
+        acc = (w_factor * (x_factor * (acc * divfactor)))
+        acc = acc.to(C.dtype.element_ty)
+
+        if has_bias:
+            bias = tl.load(bias + rn).to(C.dtype.element_ty)
+            acc = acc + bias[None, :]
+
+        C = C + (rm[:, None] * stride_cm + rn[None, :] * stride_cn)
+        mask = (rm < M)[:, None] & (rn < N)[None, :]
+        # handles write-back with reduction-splitting
+        if SPLIT_K == 1:
+            tl.store(C, acc, mask=mask)
+        else:
+            tl.atomic_add(C, acc, mask=mask)
+
+
+    def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias):
+        divfactor = 1. / (127. * 127.)
+
+        has_bias = 0 if bias is None else 1
+
+        device = a.device
+        # handle non-contiguous inputs if necessary
+        if a.stride(0) > 1 and a.stride(1) > 1:
+            a = a.contiguous()
+        if b.stride(0) > 1 and b.stride(1) > 1:
+            b = b.contiguous()
+        # checks constraints
+        assert a.shape[1] == b.shape[0], "incompatible dimensions"
+        M, K = a.shape
+        _, N = b.shape
+        # allocates output
+        c = torch.empty((M, N), device=device, dtype=torch.float16)
+        # accumulator types
+        ACC_TYPE = tl.float32 #if a.dtype in [torch.float16, torch.bfloat16, torch.float32] else tl.int32
+        # launch int8_matmul_rowwise_dequantize kernel
+        grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']), META['SPLIT_K'])
+        _int8_matmul_rowwise_dequantize[grid](a, b, c, bias, state_x, state_w, M, N, K, divfactor, has_bias,
+                        a.stride(0), a.stride(1),
+                        b.stride(0), b.stride(1),
+                        c.stride(0), c.stride(1),
+                        GROUP_M=8, ACC_TYPE=ACC_TYPE)
+        return c

bitsandbytes/triton/quantize_columnwise_and_transpose.py

+import math
+import torch
+import time
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def quantize_columnwise_and_transpose(x: torch.Tensor): return None
+else:
+
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+    # This kernel does fused columnwise quantization and transpose.
+
+    # TODO: autotune this better.
+    @triton.autotune(
+            configs=[
+                triton.Config({}, num_stages=1),
+                triton.Config({}, num_stages=2),
+                triton.Config({}, num_stages=4),
+                triton.Config({}, num_stages=8),
+                triton.Config({}, num_stages=16),
+                triton.Config({}, num_stages=1, num_warps=8),
+                triton.Config({}, num_stages=2, num_warps=8),
+                triton.Config({}, num_stages=4, num_warps=8),
+                triton.Config({}, num_stages=8, num_warps=8),
+                triton.Config({}, num_stages=16, num_warps=8),
+                triton.Config({}, num_warps=1),
+                triton.Config({}, num_warps=2),
+                triton.Config({}, num_warps=4),
+                triton.Config({}, num_warps=8),
+            ],
+            key=['n_elements']
+    )
+    @triton.jit
+    def _quantize_columnwise_and_transpose(
+        x_ptr,
+        output_ptr,
+        output_maxs,
+        n_elements,
+        M : tl.constexpr, N : tl.constexpr,
+        BLOCK_SIZE: tl.constexpr,
+        P2: tl.constexpr,
+    ):
+        pid = tl.program_id(axis=0)
+        block_start = pid
+        p2_arange = tl.arange(0, P2)
+        p2_arange_mask = p2_arange < M
+        arange =  p2_arange * N
+        offsets = block_start + arange
+        x = tl.load(x_ptr + offsets, mask=p2_arange_mask)
+        abs_x = tl.abs(x)
+        max_val = tl.max(tl.where(p2_arange_mask, abs_x, 0), axis=0)
+        output = tl.libdevice.llrint(127. * (x / max_val))
+
+        new_start = pid * M 
+        new_offsets = new_start + p2_arange
+        tl.store(output_ptr + new_offsets, output, mask=p2_arange_mask)
+        tl.store(output_maxs + pid, max_val)
+
+    def quantize_columnwise_and_transpose(x: torch.Tensor):
+        M, N = x.shape
+        output = torch.empty(N, M, device=x.device, dtype=torch.int8)
+        output_maxs = torch.empty(x.shape[1], device=x.device, dtype=torch.float16)
+
+        P2 = int(2 ** (math.ceil(math.log2(M))))
+
+        assert x.is_cuda and output.is_cuda
+        n_elements = output.numel()
+        grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
+        _quantize_columnwise_and_transpose[grid](x, output, output_maxs, n_elements, M, N, BLOCK_SIZE=M, P2=P2)
+        return output, output_maxs
+

bitsandbytes/triton/quantize_global.py

+import math
+import torch
+import time
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def quantize_global_transpose(input): return None
+    def quantize_global(x: torch.Tensor): return None
+else:
+
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+    # global quantize
+    @triton.autotune(
+            configs=[
+                triton.Config({'BLOCK_SIZE': 1024,}, num_warps=4),
+                triton.Config({'BLOCK_SIZE': 2048,}, num_stages=1),
+
+            ],
+            key=['n_elements']
+    )
+    @triton.jit
+    def _quantize_global(
+        x_ptr,
+        absmax_inv_ptr,
+        output_ptr,
+        n_elements,
+        BLOCK_SIZE: tl.constexpr,
+    ):
+        pid = tl.program_id(axis=0)
+        block_start = pid * BLOCK_SIZE
+        offsets = block_start + tl.arange(0, BLOCK_SIZE)
+        mask = offsets < n_elements
+        x = tl.load(x_ptr + offsets, mask=mask)
+        absmax_inv = tl.load(absmax_inv_ptr)
+        output = tl.libdevice.llrint(127. * (x * absmax_inv))
+        tl.store(output_ptr + offsets, output, mask=mask)
+
+    def quantize_global(x: torch.Tensor):
+        absmax = x.abs().max().unsqueeze(0)
+        absmax_inv = 1./ absmax
+        output = torch.empty(*x.shape, device='cuda', dtype=torch.int8)
+        assert x.is_cuda and output.is_cuda
+        n_elements = output.numel()
+        grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
+        _quantize_global[grid](x, absmax_inv, output, n_elements)
+        return output, absmax
+
+
+    # global quantize and transpose
+    @triton.autotune(
+            configs=[
+                triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'GROUP_M': 8}, num_warps=4),
+                triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'GROUP_M': 8}, num_warps=4),
+
+                # ...
+            ],
+            key=['M', 'N']
+    )
+    @triton.jit
+    def _quantize_global_transpose(A, absmax_inv_ptr, B, stride_am, stride_an, stride_bn, stride_bm, M, N, 
+                          BLOCK_M : tl.constexpr, 
+                          BLOCK_N : tl.constexpr, 
+                          GROUP_M : tl.constexpr):
+        pid = tl.program_id(0)
+        grid_m = (M + BLOCK_M - 1) // BLOCK_M
+        grid_n = (N + BLOCK_N - 1) // BLOCK_N
+        
+        width = GROUP_M * grid_n
+        group_id = pid // width
+        group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
+        pid_m = group_id * GROUP_M + (pid % group_size)
+        pid_n = (pid % width) // group_size
+        
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+        A = A + (rm[:, None] * stride_am + rn[None, :] * stride_an)
+        mask = (rm < M)[:, None] & (rn < N)[None, :]
+        a = tl.load(A, mask=mask)
+        absmax_inv = tl.load(absmax_inv_ptr)
+        
+        # rematerialize to save registers
+        rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+        B = B + (rm[:, None] * stride_bm + rn[None, :] * stride_bn)
+        mask = (rm < M)[:, None] & (rn < N)[None, :]
+
+        output = tl.libdevice.llrint(127. * (a * absmax_inv))
+
+        tl.store(B, output, mask=mask)
+
+    def quantize_global_transpose(input):
+        absmax = input.abs().max().unsqueeze(0)
+        absmax_inv = 1./ absmax
+        M, N = input.shape
+        out = torch.empty(N, M, device='cuda', dtype=torch.int8)
+        
+        assert out.size(0) == N and out.size(1) == M
+        assert input.stride(0) == 1 or input.stride(1) == 1
+        assert out.stride(0) == 1 or out.stride(1) == 1
+        
+        grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']),)
+        _quantize_global_transpose[grid](input, absmax_inv, out, input.stride(0), input.stride(1), out.stride(0), out.stride(1), M, N)
+        return out, absmax
+

bitsandbytes/triton/quantize_rowwise.py

+import math
+import torch
+import time
+
+from bitsandbytes.triton.triton_utils import is_triton_available
+
+if not is_triton_available():
+    def quantize_rowwise(x: torch.Tensor): return None
+else:
+
+    import triton
+    import triton.language as tl
+    from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+
+    # rowwise quantize
+
+    # TODO: autotune this better.
+    @triton.autotune(
+            configs=[
+                triton.Config({}, num_stages=1, num_warps=8),
+                triton.Config({}, num_stages=2, num_warps=8),
+                triton.Config({}, num_stages=4, num_warps=8),
+                triton.Config({}, num_stages=8, num_warps=8),
+                triton.Config({}, num_stages=1),
+                triton.Config({}, num_stages=2),
+                triton.Config({}, num_stages=4),
+                triton.Config({}, num_stages=8),
+                triton.Config({}, num_warps=1),
+                triton.Config({}, num_warps=2),
+                triton.Config({}, num_warps=4),
+                triton.Config({}, num_warps=8),
+            ],
+            key=['n_elements']
+    )
+    @triton.jit
+    def _quantize_rowwise(
+        x_ptr,
+        output_ptr,
+        output_maxs,
+        n_elements,
+        BLOCK_SIZE: tl.constexpr,
+        P2: tl.constexpr,
+    ):
+        pid = tl.program_id(axis=0)
+        block_start = pid * BLOCK_SIZE
+        arange = tl.arange(0, P2)
+        offsets = block_start + arange
+        row_mask = arange < BLOCK_SIZE
+        x = tl.load(x_ptr + offsets, mask=row_mask)
+        
+        abs_x = tl.abs(x)
+        max_val = tl.max(tl.where(row_mask, abs_x, 0), axis=0)
+        output = tl.libdevice.llrint(127. * (x / max_val))
+        tl.store(output_ptr + offsets, output, mask=row_mask)
+        tl.store(output_maxs + pid, max_val)
+
+    def quantize_rowwise(x: torch.Tensor):
+        output = torch.empty(*x.shape, device=x.device, dtype=torch.int8)
+        output_maxs = torch.empty(x.shape[0], device=x.device, dtype=torch.float16)
+
+        P2 = int(2 ** (math.ceil(math.log2(x.shape[1]))))
+
+        assert x.is_cuda and output.is_cuda
+        n_elements = output.numel()
+        grid = lambda meta: (x.shape[0],)
+        _quantize_rowwise[grid](x, output, output_maxs, n_elements, BLOCK_SIZE=x.shape[1], P2=P2)
+        return output, output_maxs
+

bitsandbytes/triton/triton_utils.py

+import importlib
+
+def is_triton_available():
+    return importlib.util.find_spec("triton") is not None

bitsandbytes/utils.py

 import shlex
 import subprocess
+import torch
 from typing import Tuple
 
+def outlier_hook(module, input):
+    assert isinstance(module, torch.nn.Linear)
+    tracer = OutlierTracer.get_instance()
+    hvalue = tracer.get_hvalue(module.weight)
+    if hvalue not in tracer.hvalue2outlier_idx:
+        outlier_idx = find_outlier_dims(module.weight)
+        tracer.outliers.append(outlier_idx)
+        tracer.hvalues.append(hvalue)
+        if len(tracer.outliers) > 1:
+            # assign the current layer the outlier idx found from the weight
+            # of the previous linear layer
+            if tracer.outliers[-1].numel() > 0:
+                assert tracer.outliers[-1].max() < module.weight.shape[1]
+            tracer.hvalue2outlier_idx[hvalue] = tracer.outliers[-1]
+
+        else:
+            # first layer, we cannot use the weight for outlier detection
+            # we follow a mixed approach:
+            # (1) zscore test of std of hidden dimension
+            # (2) magnitude > 6 test
+            merged = input[0].view(-1, input[0].shape[-1])
+            # (1) zscore test of std of hidden dimension
+            outlier_idx = find_outlier_dims(merged, reduction_dim=1, zscore=3)
+            # (2) magnitude > 6 test
+            dims = (torch.abs(input[0])> 6).sum(dim=list(range(len(input[0].shape)-1)))
+            outlier_idx2 = torch.where(dims > 0)[0]
+            outlier_idx = torch.cat([outlier_idx, outlier_idx2]).unique()
+            tracer.hvalue2outlier_idx[hvalue] = outlier_idx
+    else:
+        for hook in tracer.hooks:
+            hook.remove()
+
+
+class OutlierTracer(object):
+    _instance = None
+
+    def __init__(self):
+        raise RuntimeError("Call get_instance() instead")
+
+    def initialize(self, model):
+        self.last_w = None
+        self.current_outlier_dims = None
+        self.hvalues = []
+        self.outliers = []
+        self.hvalue2outlier_idx = {}
+        self.initialized = True
+        self.hooks = []
+
+        for n, m in model.named_modules():
+            if isinstance(m, torch.nn.Linear):
+                self.hooks.append(m.register_forward_pre_hook(outlier_hook))
+
+    def is_initialized(self):
+        return getattr(self, 'initialized', False)
+
+    def get_hvalue(self, weight):
+        return weight.data.storage().data_ptr()
+
+    def get_outliers(self, weight):
+        if not self.is_initialized():
+            print('Outlier tracer is not initialized...')
+            return None
+        hvalue = self.get_hvalue(weight)
+        if hvalue in self.hvalue2outlier_idx:
+            return self.hvalue2outlier_idx[hvalue]
+        else:
+            return None
+
+    @classmethod
+    def get_instance(cls):
+        if cls._instance is None:
+            cls._instance = cls.__new__(cls)
+        return cls._instance
+
+def find_outlier_dims(weight, reduction_dim=0, zscore=4.0, topk=None, rdm=False):
+    if rdm:
+        return torch.randint(0, weight.shape[1], size=(topk,), device=weight.device).long()
+
+    m = weight.mean(reduction_dim)
+    mm = m.mean()
+    mstd = m.std()
+    zm = (m-mm)/mstd
+
+    std = weight.std(reduction_dim)
+    stdm = std.mean()
+    stdstd = std.std()
+
+    zstd = (std-stdm)/stdstd
+
+    if topk is not None:
+        val, idx = torch.topk(std.abs(), k=topk, dim=0)
+    else:
+        idx = torch.where(zstd > zscore)[0]
+
+    return idx
+
+def replace_linear(model, linear_replacement, skip_modules=["lm_head"], copy_weights=False, post_processing_function=None):
+    """
+    Replace linear modules with a new Linear module.
+
+    Parameters:
+        model (`torch.nn.Module`):
+            Input model or `torch.nn.Module` as the function is run recursively.
+        linear_replacement (`torch.nn.Module`):
+            The linear module that replaces the old one. Only expects standard arguments.
+            If other arguments need to be passed, use a lambda.
+        skip_modules (`List[str]`, *optional*, defaults to `lm_head`):
+            List of modules names not to convert. Defaults to `lm_head`.
+        copy_weights (`bool`):
+            Copy the weights from the old linear module to the new one
+        post_processing_fun_name (`str`):
+            A function name of the replacement linear class that is called
+            after processing.
+    """
+    for name, module in model.named_children():
+        if len(list(module.children())) > 0:
+            replace_linear(module, linear_replacement, skip_modules, copy_weights, post_processing_function)
+
+        if isinstance(module, torch.nn.Linear) and name not in skip_modules:
+            old_module = model._modules[name]
+            model._modules[name] = linear_replacement(
+                module.in_features,
+                module.out_features,
+                module.bias is not None,
+            )
+            if copy_weights:
+                model._modules[name].weight = old_module.weight
+                model._modules[name].bias = old_module.bias
+
+            if post_processing_function is not None:
+               func = getattr(module, post_processing_function, None)
+               if func is not None: func(module)
+    return model
+
+
 
 def execute_and_return(command_string: str) -> Tuple[str, str]:
     def _decode(subprocess_err_out_tuple):

compile_from_source.md

 # Compiling from source
 
 Basic steps.
-1. `make [target]` where `[target]` is among `cuda92, cuda10x, cuda110, cuda11x, cpuonly`
-2. `CUDA_VERSION=XXX python setup.py install`
+1. `CUDA_VERSION=XXX make [target]` where `[target]` is among `cuda92, cuda10x, cuda110, cuda11x, cuda12x, cpuonly`
+2. `python setup.py install`
 
 To run these steps you will need to have the nvcc compiler installed that comes with a CUDA installation. If you use anaconda (recommended) then you can figure out which version of CUDA you are using with PyTorch via the command `conda list | grep cudatoolkit`. Then you can install the nvcc compiler by downloading and installing the same CUDA version from the [CUDA toolkit archive](https://developer.nvidia.com/cuda-toolkit-archive).
 
-For your convenience, there is an installation script in the root directory that installs CUDA 11.1 locally and configures it automatically. After installing you should add the `bin` sub-directory to the `$PATH` variable to make the compiler visible to your system. To do this you can add this to your `.bashrc` by executing these commands:
+You can install CUDA locally without sudo by following the following steps:
+
 ```bash
-echo "export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64/" >> ~/.bashrc
-echo "export PATH=$PATH:/usr/local/cuda/bin/" >> ~/.bashrc
-source ~/.bashrc
+wget https://raw.githubusercontent.com/TimDettmers/bitsandbytes/main/cuda_install.sh
+# Syntax cuda_install CUDA_VERSION INSTALL_PREFIX EXPORT_TO_BASH
+#   CUDA_VERSION in {110, 111, 112, 113, 114, 115, 116, 117, 118, 120, 121}
+#   EXPORT_TO_BASH in {0, 1} with 0=False and 1=True 
+
+# For example, the following installs CUDA 11.7 to ~/local/cuda-11.7 and exports the path to your .bashrc
+bash cuda install 117 ~/local 1 
 ```
 
 By default, the Makefile will look at your `CUDA_HOME` environmental variable to find your CUDA version for compiling the library. If this path is not set it is inferred from the path of your `nvcc` compiler.
 
 Either `nvcc` needs to be in path for the `CUDA_HOME` variable needs to be set to the CUDA directory root (e.g. `/usr/local/cuda`) in order for compilation to succeed
 
+If you type `nvcc` and it cannot be found, you might need to add to your path or set the CUDA_HOME variable. You can run `python -m bitsandbytes` to find the path to CUDA. For example if `python -m bitsandbytes` shows you the following:
+```
+++++++++++++++++++ /usr/local CUDA PATHS +++++++++++++++++++
+/usr/local/cuda-11.7/targets/x86_64-linux/lib/libcudart.so
+```
+You can set `CUDA_HOME` to `/usr/local/cuda-11.7`. For example, you might be able to compile like this.
+
+``CUDA_HOME=~/local/cuda-11.7 CUDA_VERSION=117 make cuda11x``
+
+
 If you have problems compiling the library with these instructions from source, please open an issue.

csrc/kernels.cu

@@ -329,6 +329,13 @@ __device__ unsigned char dQuantizeNF4(float x)
         else
           return 0b0000;
 }
+// sign function for lion
+// taken from https://stackoverflow.com/a/4609795, but not sure if there's a proper way to do this in CUDA
+
+template <typename T> __device__ int sgn(T val)
+{
+  return (T(0) < val) - (val < T(0));
+}
 
 template <int STOCHASTIC>
 __device__ unsigned char dQuantize(float* smem_code, const float rand, float x)
@@ -857,7 +864,6 @@ __global__ void kDequantizeBlockwise(float *code, unsigned char * A, float * abs
     __syncthreads();
     LoadChar(loadchar).Load(&(A[i]), qvals, valid_items_load, 128);
 
-
     switch(DATA_TYPE)
     {
         case General8bit:
@@ -1081,7 +1087,7 @@ template<typename T, int OPTIMIZER, int BLOCK_SIZE, int NUM_VALS>
 __launch_bounds__(BLOCK_SIZE/NUM_VALS, 1)
 __global__ void kPreconditionOptimizer32bit1State(T* g, T* p,
                 float* state1, float *unorm,
-                const float beta1, const float eps, const float weight_decay,
+                const float beta1, const float beta2, const float eps, const float weight_decay,
                 const int step, const float lr, const float gnorm_scale, const int n)
 {
 
@@ -1128,6 +1134,9 @@ __global__ void kPreconditionOptimizer32bit1State(T* g, T* p,
                     s1_vals[j] = s1_vals[j]*beta1 + ((float)g_vals[j]); // state update
                   s1_vals[j] = s1_vals[j]*s1_vals[j]; // update norm
                   break;
+              case LION:
+                  s1_vals[j] = s1_vals[j]*beta2 + ((1.0f-beta2)*(float)g_vals[j]); // state update
+                  break;
               case RMSPROP:
                   s1_vals[j] = s1_vals[j]*beta1 + ((1.0f-beta1)*((float)g_vals[j])*((float)g_vals[j])); // state update
                   s1_vals[j] = __fdividef((float)g_vals[j],sqrtf(s1_vals[j])+eps); // update value
@@ -1159,7 +1168,7 @@ template<typename T, int OPTIMIZER>
 __launch_bounds__(TH, 1)
 __global__ void kOptimizer32bit1State(T *g, T *p,
                 float *state1, float *unorm, const float max_unorm, const float param_norm,
-                const float beta1, const float eps, const float weight_decay,
+                const float beta1, const float beta2, const float eps, const float weight_decay,
                 const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
@@ -1228,6 +1237,10 @@ __global__ void kOptimizer32bit1State(T *g, T *p,
 
 										p_vals[j] = ((float)p_vals[j]) + update_scale*(-lr*(s1_vals[j]));
 										break;
+								case LION:
+										p_vals[j] = ((float)p_vals[j]) - update_scale*(lr*sgn(((float)s1_vals[j])*beta1 + ((1.0f-beta1)*((float)g_vals[j]))));
+										s1_vals[j] = s1_vals[j]*beta2 + ((1.0f-beta2)*((float)g_vals[j]));
+										break;
 								case RMSPROP:
 										s1_vals[j] = s1_vals[j]*beta1 + ((1.0f-beta1)*((float)g_vals[j])*((float)g_vals[j]));
 										p_vals[j] = ((float)p_vals[j]) - update_scale*(lr*__fdividef((float)g_vals[j],sqrtf((float)s1_vals[j])+eps));
@@ -1496,7 +1509,7 @@ __global__ void
 __launch_bounds__(NUM_THREADS, 2)
 kPreconditionOptimizerStatic8bit1State(T* p, T* __restrict__ const g, unsigned char*__restrict__  const state1,
                 float *unorm,
-                const float beta1,
+                const float beta1, const float beta2,
                 const float eps, const int step,
                 float* __restrict__ const quantiles1,
                 float* max1, float* new_max1,
@@ -1557,6 +1570,9 @@ kPreconditionOptimizerStatic8bit1State(T* p, T* __restrict__ const g, unsigned c
                     if(unorm != NULL)
                       local_unorm += s1_vals[j]*s1_vals[j];
                     break;
+              case LION:
+                  s1_vals[j] = s1_vals[j]*beta2 + ((1.0f-beta2)*g_val);
+                  break;
               case RMSPROP:
                     s1_vals[j] = s1_vals[j]*beta1 + ((1.0f-beta1)*(g_val*g_val));
                   break;
@@ -1580,9 +1596,10 @@ kPreconditionOptimizerStatic8bit1State(T* p, T* __restrict__ const g, unsigned c
 
 template<typename T, int OPTIMIZER>
 __global__ void
+__launch_bounds__(1024, 1)
 kOptimizerStatic8bit1State(T* p, T* const g, unsigned char* state1,
                 const float *unorm, const float max_unorm, const float param_norm,
-                const float beta1,
+                const float beta1, const float beta2,
                 const float eps, const int step, const float lr,
                 float* __restrict__ const quantiles1,
                 float* max1, float* new_max1,
@@ -1645,8 +1662,19 @@ kOptimizerStatic8bit1State(T* p, T* const g, unsigned char* state1,
         {
             g_val = float(g_vals[j]);
             g_val *= gnorm_scale;
-            if(weight_decay > 0.0f)
+
+            if(weight_decay > 0.0f) {
+              switch(OPTIMIZER) {
+                case MOMENTUM:
+                case RMSPROP:
                   g_val += ((float)p_vals[j])*weight_decay;
+                  break;
+                case LION:
+                  p_vals[j] = ((float)p_vals[j])*(1.0f-lr*weight_decay);
+                  break;
+              }
+            }
+
             s1_vals[j] = smem_quantiles1[c1s[j]]*max1[0];
 
             switch(OPTIMIZER)
@@ -1659,6 +1687,10 @@ kOptimizerStatic8bit1State(T* p, T* const g, unsigned char* state1,
 
                   p_vals[j] = ((float)p_vals[j]) + (-lr*update_scale*(s1_vals[j]));
                   break;
+              case LION:
+                  p_vals[j] = ((float)p_vals[j]) - (lr*sgn(((float)s1_vals[j])*beta1 + ((1.0f-beta1)*((float)g_val))));
+                  s1_vals[j] = s1_vals[j]*beta2 + ((1.0f-beta2)*g_val);
+                  break;
               case RMSPROP:
                   s1_vals[j] = s1_vals[j]*beta1 + ((1.0f-beta1)*(g_val*g_val));
                   p_vals[j] = ((float)p_vals[j]) - (lr*__fdividef(g_val,sqrtf(s1_vals[j])+eps));
@@ -1999,8 +2031,18 @@ kOptimizerStatic8bit1StateBlockwise(T* p, T* __restrict__ const g, unsigned char
             g_val *= gnorm_scale;
             if(!skip_zeros || (skip_zeros && ((float)g_vals[j] != 0.0f)))
             {
-							if(weight_decay > 0.0f)
+              if(weight_decay > 0.0f) {
+                switch(OPTIMIZER) {
+                  case MOMENTUM:
+                  case ADAGRAD:
+                  case RMSPROP:
                     g_val += ((float)p_vals[j])*weight_decay;
+                    break;
+                  case LION:
+                    p_vals[j] = ((float)p_vals[j])*(1.0f-lr*weight_decay);
+                    break;
+                }
+              }
 
 							s1_vals[j] = smem_quantiles1[lane_id][c1s[j]]*absmax1[i/BLOCK_SIZE];
 
@@ -2012,6 +2054,11 @@ kOptimizerStatic8bit1StateBlockwise(T* p, T* __restrict__ const g, unsigned char
 										else
 											s1_vals[j] = (s1_vals[j]*beta1) + g_val;
 										break;
+									case LION:
+										// here, using gvals[j] to store the gradient smoothed by beta1 for the following parameter update, before the momentum is updated by beta2
+										g_vals[j] = lr*sgn(((float)s1_vals[j])*beta1 + ((1.0f-beta1)*g_val));
+										s1_vals[j] = s1_vals[j]*beta2 + ((1.0f-beta2)*g_val);
+										break;
 									case RMSPROP:
 										s1_vals[j] = s1_vals[j]*beta1 + ((1.0f-beta1)*(g_val*g_val));
 										break;
@@ -2049,6 +2096,9 @@ kOptimizerStatic8bit1StateBlockwise(T* p, T* __restrict__ const g, unsigned char
 									case MOMENTUM:
 										p_vals[j] = ((float)p_vals[j]) - lr*(s1_vals[j]);
 										break;
+									case LION:
+										p_vals[j] = ((float)p_vals[j]) - ((float)g_vals[j]);
+										break;
 									case RMSPROP:
 										g_val = g_vals[j];
 										p_vals[j] = ((float)p_vals[j]) - lr*(__fdividef(g_val, sqrtf(s1_vals[j])+eps));
@@ -3607,24 +3657,28 @@ template __global__ void kEstimateQuantiles(half *__restrict__ const A, float *c
 #define MAKE_PreconditionOptimizer32bit1State(oname, gtype) \
 template __global__ void kPreconditionOptimizer32bit1State<gtype, oname, 4096, 8>(gtype* g, gtype* p, \
                 float* state1, float *unorm, \
-                const float beta1, const float eps, const float weight_decay, \
+                const float beta1, const float beta2, const float eps, const float weight_decay, \
                 const int step, const float lr, const float gnorm_scale, const int n); \
 
 MAKE_PreconditionOptimizer32bit1State(MOMENTUM, half)
 MAKE_PreconditionOptimizer32bit1State(MOMENTUM, float)
 MAKE_PreconditionOptimizer32bit1State(RMSPROP, half)
 MAKE_PreconditionOptimizer32bit1State(RMSPROP, float)
+MAKE_PreconditionOptimizer32bit1State(LION, half)
+MAKE_PreconditionOptimizer32bit1State(LION, float)
 MAKE_PreconditionOptimizer32bit1State(ADAGRAD, half)
 MAKE_PreconditionOptimizer32bit1State(ADAGRAD, float)
 
 #define MAKE_Optimizer32bit1State(oname, gtype) \
 template __global__ void kOptimizer32bit1State<gtype, oname>(gtype* g, gtype* p, float* state1, float *unorm, const float max_unorm, const float param_norm, \
-    const float beta1, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n); \
+    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_Optimizer32bit1State(MOMENTUM, half)
 MAKE_Optimizer32bit1State(MOMENTUM, float)
 MAKE_Optimizer32bit1State(RMSPROP, half)
 MAKE_Optimizer32bit1State(RMSPROP, float)
+MAKE_Optimizer32bit1State(LION, half)
+MAKE_Optimizer32bit1State(LION, float)
 MAKE_Optimizer32bit1State(ADAGRAD, half)
 MAKE_Optimizer32bit1State(ADAGRAD, float)
 
@@ -3649,6 +3703,7 @@ template __global__ void kOptimizer32bit2State<__nv_bfloat16, ADAM>(__nv_bfloat1
 template __global__ void kPreconditionOptimizerStatic8bit1State<gtype, oname>(gtype* p, gtype* __restrict__ const g, unsigned char*__restrict__  const state1,  \
                 float *unorm,  \
                 const float beta1,  \
+                const float beta2,  \
                 const float eps, const int step,  \
                 float* __restrict__ const quantiles1,  \
                 float* max1, float* new_max1,  \
@@ -3660,11 +3715,14 @@ MAKE_PreconditionStatic8bit1State(MOMENTUM, half)
 MAKE_PreconditionStatic8bit1State(MOMENTUM, float)
 MAKE_PreconditionStatic8bit1State(RMSPROP, half)
 MAKE_PreconditionStatic8bit1State(RMSPROP, float)
+MAKE_PreconditionStatic8bit1State(LION, half)
+MAKE_PreconditionStatic8bit1State(LION, float)
 
 #define MAKE_optimizerStatic8bit1State(oname, gtype) \
 template __global__ void kOptimizerStatic8bit1State<gtype, oname>(gtype* p, gtype* const g, unsigned char* state1,  \
                 const float *unorm, const float max_unorm, const float param_norm, \
                 const float beta1,  \
+                const float beta2,  \
                 const float eps, const int step, const float lr, \
                 float* __restrict__ const quantiles1,  \
                 float* max1, float* new_max1,  \
@@ -3676,6 +3734,8 @@ MAKE_optimizerStatic8bit1State(MOMENTUM, half)
 MAKE_optimizerStatic8bit1State(MOMENTUM, float)
 MAKE_optimizerStatic8bit1State(RMSPROP, half)
 MAKE_optimizerStatic8bit1State(RMSPROP, float)
+MAKE_optimizerStatic8bit1State(LION, half)
+MAKE_optimizerStatic8bit1State(LION, float)
 
 #define MAKE_PreconditionStatic8bit2State(oname, gtype) \
 template __global__ void kPreconditionOptimizerStatic8bit2State<gtype, oname>(gtype* p, gtype* __restrict__ const g, unsigned char*__restrict__  const state1, unsigned char* __restrict__ const state2, \
@@ -3762,7 +3822,6 @@ template __global__ void kDequantizeBlockwise<float, 512, 64, 8, General8bit>(fl
 template __global__ void kDequantizeBlockwise<half, 512, 64, 8, NF4>(float *code, unsigned char * A, float * absmax, half *out, const int blocksize, const int n);
 template __global__ void kDequantizeBlockwise<float, 512, 64, 8, NF4>(float *code, unsigned char * A, float * absmax, float *out, const int blocksize, const int n);
 
-
 #define MAKE_OptimizerStatic8bit2StateBlockwise(oname, gtype, block_size, num_per_thread) \
 template __global__ void kOptimizerStatic8bit2StateBlockwise<gtype, oname, block_size, num_per_thread>(gtype* p, gtype* __restrict__ const g, unsigned char* state1, unsigned char* state2, \
                 const float beta1, const float beta2, \
@@ -3791,5 +3850,7 @@ MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, float, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, half, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(RMSPROP, float, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(RMSPROP, half, 2048, 8)
+MAKE_OptimizerStatic8bit1StateBlockwise(LION, float, 2048, 8)
+MAKE_OptimizerStatic8bit1StateBlockwise(LION, half, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(ADAGRAD, float, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(ADAGRAD, half, 2048, 8)

csrc/kernels.cuh

@@ -34,20 +34,20 @@ __global__ void kOptimizer32bit2State(T* g, T* p,
 template<typename T, int OPTIMIZER, int BLOCK_SIZE, int NUM_VALS>
 __global__ void kPreconditionOptimizer32bit1State(T* g, T* p,
                 float* state1, float *unorm,
-                const float beta1, const float eps, const float weight_decay,
+                const float beta1, const float beta2, const float eps, const float weight_decay,
                 const int step, const float lr, const float gnorm_scale, const int n);
 
 template<typename T, int OPTIMIZER>
 __global__ void kOptimizer32bit1State(T* g, T* p,
                 float* state1,  float *unorm, const float max_unorm, const float param_norm,
-                const float beta1, const float eps, const float weight_decay,
+                const float beta1, const float beta2, const float eps, const float weight_decay,
                 const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER>
 __global__ void
 kPreconditionOptimizerStatic8bit1State(T* p, T* __restrict__ const g, unsigned char*__restrict__  const state1,
                 float *unorm,
-                const float beta1,
+                const float beta1, const float beta2,
                 const float eps, const int step,
                 float* __restrict__ const quantiles1,
                 float* max1, float* new_max1,
@@ -59,7 +59,7 @@ template<typename T, int OPTIMIZER>
 __global__ void
 kOptimizerStatic8bit1State(T* p, T* const g, unsigned char* state1,
                 const float *unorm, const float max_unorm, const float param_norm,
-                const float beta1,
+                const float beta1, const float beta2,
                 const float eps, const int step, const float lr,
                 float* __restrict__ const quantiles1,
                 float* max1, float* new_max1,

csrc/ops.cu

@@ -54,8 +54,6 @@ template <typename T, int STOCHASTIC, int DATA_TYPE> void quantizeBlockwise(floa
 {
   int num_blocks = n/blocksize;
   num_blocks = n % blocksize == 0 ? num_blocks : num_blocks + 1;
-  if(STOCHASTIC == 1)
-    assert(blocksize == 4096);
 
   if(blocksize == 4096)
     kQuantizeBlockwise<T, 4096, 4, STOCHASTIC, 0><<<num_blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
@@ -121,17 +119,28 @@ template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
 		case MOMENTUM:
     case RMSPROP:
     case ADAGRAD:
-
       if(max_unorm > 0.0f)
 			{
 				CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
-				kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<num_blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
+				kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<num_blocks, 512>>>(g, p, state1, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
         CUDA_CHECK_RETURN(cudaPeekAtLastError());
 			}
 
-			kOptimizer32bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+			kOptimizer32bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
       CUDA_CHECK_RETURN(cudaPeekAtLastError());
 			break;
+    case LION:
+      // in lion, the momentum update after the parameter update
+      kOptimizer32bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+      CUDA_CHECK_RETURN(cudaPeekAtLastError());
+
+      if(max_unorm > 0.0f)
+      {
+        CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
+        kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<num_blocks, 512>>>(g, p, state1, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
+        CUDA_CHECK_RETURN(cudaPeekAtLastError());
+      }
+      break;
 	}
 }
 
@@ -165,11 +174,21 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
     case RMSPROP:
     case ADAGRAD:
 			CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
-			kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+			kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 256>>>(p, g, state1, unorm, beta1, beta2, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			kOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
+																														quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
 			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			kOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
+			break;
+    case LION:
+      // in lion, the momentum update happens after the parameter update
+      kOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
                                                             quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
       CUDA_CHECK_RETURN(cudaPeekAtLastError());
+
+      CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
+      kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<num_blocks, 256>>>(p, g, state1, unorm, beta1, beta2, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+      CUDA_CHECK_RETURN(cudaPeekAtLastError());
       break;
 		default:
 			break;
@@ -199,6 +218,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 		case MOMENTUM:
 		case RMSPROP:
     case ADAGRAD:
+    case LION:
 			num_blocks = n/BLOCKSIZE_1STATE;
 			num_blocks = n % BLOCKSIZE_1STATE == 0 ? num_blocks : num_blocks + 1;
 			kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<num_blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
@@ -780,6 +800,8 @@ MAKE_optimizer32bit(MOMENTUM, half)
 MAKE_optimizer32bit(MOMENTUM, float)
 MAKE_optimizer32bit(RMSPROP, half)
 MAKE_optimizer32bit(RMSPROP, float)
+MAKE_optimizer32bit(LION, half)
+MAKE_optimizer32bit(LION, float)
 MAKE_optimizer32bit(ADAGRAD, half)
 MAKE_optimizer32bit(ADAGRAD, float)
 
@@ -799,6 +821,8 @@ MAKE_optimizerStatic8bit(MOMENTUM, half)
 MAKE_optimizerStatic8bit(MOMENTUM, float)
 MAKE_optimizerStatic8bit(RMSPROP, half)
 MAKE_optimizerStatic8bit(RMSPROP, float)
+MAKE_optimizerStatic8bit(LION, half)
+MAKE_optimizerStatic8bit(LION, float)
 
 #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
 template void optimizerStatic8bitBlockwise<gtype, optim_name>(gtype* p, gtype* g, \
@@ -811,6 +835,8 @@ MAKE_optimizerStatic8bitBlockwise(half, MOMENTUM);
 MAKE_optimizerStatic8bitBlockwise(float, MOMENTUM);
 MAKE_optimizerStatic8bitBlockwise(half, RMSPROP);
 MAKE_optimizerStatic8bitBlockwise(float, RMSPROP);
+MAKE_optimizerStatic8bitBlockwise(half, LION);
+MAKE_optimizerStatic8bitBlockwise(float, LION);
 MAKE_optimizerStatic8bitBlockwise(half, ADAGRAD);
 MAKE_optimizerStatic8bitBlockwise(float, ADAGRAD);
 

csrc/ops.cuh

@@ -75,6 +75,7 @@ typedef enum Optimizer_t
   RMSPROP = 2,
   LARS = 3,
   ADAGRAD = 4,
+  LION = 5,
 } Optimizer_t;
 
 typedef enum Transform_t

csrc/pythonInterface.c

@@ -38,7 +38,7 @@ MAKE_ELEMENTWISE_FUNC(_mul, fp32, float, _MUL)
 
 
 #define MAKE_FUNC32(fname, oname, gtype, gbits) \
-void fname##32bit_g##gbits(gtype *g, gtype *p, \
+void fname##32bit_grad_##gbits(gtype *g, gtype *p, \
                float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
                const float beta1, const float beta2, const float eps, const float weight_decay, \
                const int step, const float lr, float gnorm_scale, bool skip_zeros, const int n) \
@@ -51,11 +51,13 @@ MAKE_FUNC32(adam, ADAM, half, fp16)
 MAKE_FUNC32(adam, ADAM, __nv_bfloat16, bf16)
 MAKE_FUNC32(rmsprop, RMSPROP, float, 32)
 MAKE_FUNC32(rmsprop, RMSPROP, half, 16)
+MAKE_FUNC32(lion, LION, float, 32)
+MAKE_FUNC32(lion, LION, half, 16)
 MAKE_FUNC32(adagrad, ADAGRAD, float, 32)
 MAKE_FUNC32(adagrad, ADAGRAD, half, 16)
 
 #define MAKE_FUNC8(fname, oname, gtype, gbits) \
-void fname##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
+void fname##_static_8bit_grad_##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
 								float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
@@ -73,9 +75,11 @@ MAKE_FUNC8(momentum, MOMENTUM, float, 32)
 MAKE_FUNC8(momentum, MOMENTUM, half, 16)
 MAKE_FUNC8(rmsprop, RMSPROP, float, 32)
 MAKE_FUNC8(rmsprop, RMSPROP, half, 16)
+MAKE_FUNC8(lion, LION, float, 32)
+MAKE_FUNC8(lion, LION, half, 16)
 
 #define MAKE_BLOCKWISE8(fname, optim_name, gtype, gbits) \
-void fname##_8bit_blockwise_##gbits(gtype* p, gtype* g, \
+void fname##_8bit_blockwise_grad_##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)\
 {	optimizerStatic8bitBlockwise<gtype, optim_name>(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); }\
@@ -89,6 +93,8 @@ MAKE_BLOCKWISE8(rmsprop, RMSPROP, float, fp32)
 MAKE_BLOCKWISE8(adagrad, ADAGRAD, half, fp16)
 MAKE_BLOCKWISE8(adagrad, ADAGRAD, float, fp32)
 MAKE_BLOCKWISE8(adam, ADAM, __nv_bfloat16, bf16)
+MAKE_BLOCKWISE8(lion, LION, half, fp16)
+MAKE_BLOCKWISE8(lion, LION, float, fp32)
 
 
 void percentileClipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping<float>(g, gnorm_vec, step, n); }
@@ -96,8 +102,6 @@ void percentileClipping_g16(half * g, float *gnorm_vec, int step, const int n){
 
 void quantizeBlockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<half, 0, General8bit>(code, A, absmax, out, NULL, 0, blocksize, n); }
 void quantizeBlockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<float, 0, General8bit>(code, A, absmax, out, NULL, 0, blocksize, n); }
-void quantizeBlockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<half, 1, General8bit>(code, A, absmax, out, rand, rand_offset, 4096, n); }
-void quantizeBlockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<float, 1, General8bit>(code, A, absmax, out, rand, rand_offset, 4096, n); }
 void quantizeBlockwise_fp16_fp4(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<half, 0, FP4>(NULL, A, absmax, out, NULL, 0, blocksize, n); }
 void quantizeBlockwise_fp32_fp4(float * code, float *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<float, 0, FP4>(NULL, A, absmax, out, NULL, 0, blocksize, n); }
 void quantizeBlockwise_fp16_nf4(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<half, 0, NF4>(NULL, A, absmax, out, NULL, 0, blocksize, n); }
@@ -110,6 +114,7 @@ void dequantizeBlockwise_fp32_fp4(float *code, unsigned char *A, float *absmax,
 void dequantizeBlockwise_fp16_nf4(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise<half, NF4>(NULL, A, absmax, out, blocksize, n); } \
 void dequantizeBlockwise_fp32_nf4(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise<float, NF4>(NULL, A, absmax, out, blocksize, n); }
 
+
 #define MAKE_FUNC_TRANSFORM(fbits, fsrc, ftrgt, ftranspose, dtype, src, target, transpose, bits) \
 void transform_##fbits##_##fsrc##_to_##ftrgt##_##ftranspose(cublasLtHandle_t ltHandle, dtype *A, dtype *out, int dim1, int dim2) \
 { \
@@ -169,8 +174,6 @@ extern "C"
 	void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); }
   void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
   void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
-  void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); }
-  void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); }
 
   void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
   void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
@@ -185,11 +188,11 @@ extern "C"
   void cdequantize_blockwise_fp32_nf4(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32_nf4(code, A, absmax, out, blocksize, n); }
 
 	#define MAKE_CFUNC32(name, gtype, gbits) \
-	void c##name##32bit_g##gbits(gtype *g, gtype *p, \
+	void c##name##32bit_grad_##gbits(gtype *g, gtype *p, \
 								 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
 								 const float beta1, const float beta2, const float eps, const float weight_decay, \
 								 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
-	{ name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
+	{ name##32bit_grad_##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
 	MAKE_CFUNC32(adam, float, fp32)
 	MAKE_CFUNC32(adam, half, fp16)
@@ -198,11 +201,13 @@ extern "C"
 	MAKE_CFUNC32(momentum, half, 16)
 	MAKE_CFUNC32(rmsprop, float, 32)
 	MAKE_CFUNC32(rmsprop, half, 16)
+	MAKE_CFUNC32(lion, float, 32)
+	MAKE_CFUNC32(lion, half, 16)
 	MAKE_CFUNC32(adagrad, float, 32)
 	MAKE_CFUNC32(adagrad, half, 16)
 
 	#define MAKE_CFUNC8(name, gtype, gbits) \
-	void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
+	void c##name##_static_8bit_grad_##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
                 float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
@@ -210,7 +215,7 @@ extern "C"
                 float* max1, float* max2, float* new_max1, float* new_max2, \
                 float weight_decay, float gnorm_scale, int n) \
   {  \
-	    name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
+	    name##_static_8bit_grad_##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
 			                                 quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
   } \
 
@@ -220,12 +225,14 @@ extern "C"
 	MAKE_CFUNC8(momentum, half, 16)
 	MAKE_CFUNC8(rmsprop, float, 32)
 	MAKE_CFUNC8(rmsprop, half, 16)
+	MAKE_CFUNC8(lion, float, 32)
+	MAKE_CFUNC8(lion, half, 16)
 
   #define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
-  void c##fname##_8bit_blockwise_##gbits(gtype* p, gtype* g, \
+  void c##fname##_8bit_blockwise_grad_##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n) \
-  {	fname##_8bit_blockwise_##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \
+  {	fname##_8bit_blockwise_grad_##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \
 
 	MAKE_CBLOCKWISE8(adam, ADAM, half, fp16)
 	MAKE_CBLOCKWISE8(adam, ADAM, float, fp32)
@@ -236,6 +243,8 @@ extern "C"
 	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, fp16)
 	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, fp32)
 	MAKE_CBLOCKWISE8(adam, ADAM, __nv_bfloat16, bf16)
+	MAKE_CBLOCKWISE8(lion, LION, half, fp16)
+	MAKE_CBLOCKWISE8(lion, LION, float, fp32)
 
 	void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); }
 	void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); }