bitsandbytes 0.50.0__py3-none-win_arm64.whl

bitsandbytes/backends/triton/kernels_optim.py

@@ -0,0 +1,1177 @@
+import math
+from typing import Optional
+
+import torch
+
+import triton
+import triton.language as tl
+
+# from triton.language.extra import libdevice
+from .kernels_8bit_quant import (
+    dequant_8bit_blockwise,
+    dequant_8bit_blockwise_kernel_util,
+    quantize_8bit_blockwise_kernel_util,
+    quantize_blockwise_triton,
+)
+
+MOMENTUM = 0
+RMSPROP = 1
+ADAGRAD = 2
+ADAM = 3
+# LION should be larger than MOMENTUM, RMSPROP, ADAGRAD due to comparison in kernels
+LION = 4
+ADEMAMIX = 5
+
+name2optimizer_id = {
+    "momentum": MOMENTUM,
+    "lars": MOMENTUM,
+    "rmsprop": RMSPROP,
+    "adagrad": ADAGRAD,
+    "adam": ADAM,
+    "lamb": ADAM,
+    "lion": LION,
+    "ademamix": ADEMAMIX,
+}
+
+
+@triton.jit
+def _optimizer_precondition_2state_32bit(
+    g_ptr,
+    p_ptr,
+    state1_ptr,
+    state2_ptr,
+    unorm_ptr,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    eps: tl.constexpr,
+    weight_decay: tl.constexpr,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    gnorm_scale: tl.constexpr,
+    n_elements,
+    OPTIMIZER_ID: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+):
+    """Preprocessing optimizer, computing update norm (2-state optimizer)"""
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE * N_PER_TH)
+    mask = offsets < n_elements
+
+    g_vals = tl.load(g_ptr + offsets, mask=mask, other=0.0)
+    s1_vals = tl.load(state1_ptr + offsets, mask=mask, other=0.0)
+    s2_vals = tl.load(state2_ptr + offsets, mask=mask, other=0.0)
+
+    g_vals = gnorm_scale * g_vals
+
+    correction1 = 1.0 / (1.0 - beta1_step)
+    correction2 = 1.0 / (1.0 - beta2_step)
+
+    if OPTIMIZER_ID == 3:  # ADAM
+        s1_vals = s1_vals * beta1 + (1.0 - beta1) * g_vals
+        s2_vals = s2_vals * beta2 + (1.0 - beta2) * g_vals * g_vals
+
+        s1_vals = s1_vals * correction1
+        s2_vals = s2_vals * correction2
+
+        update_vals = s1_vals / (tl.sqrt(s2_vals) + eps)
+
+        update_norm = update_vals * update_vals
+
+    elif OPTIMIZER_ID == 5:  # ADEMAMIX
+        update_norm = s1_vals
+
+    total_norm = tl.sum(tl.where(mask, update_norm, 0.0))
+
+    tl.atomic_add(unorm_ptr, total_norm)
+
+
+@triton.jit
+def _optimizer_precondition_1state_32bit(
+    g_ptr,
+    p_ptr,
+    state1_ptr,
+    state2_ptr,
+    unorm_ptr,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    eps: tl.constexpr,
+    weight_decay,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    gnorm_scale: tl.constexpr,
+    n_elements,
+    OPTIMIZER_ID: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+):
+    """Preprocessing optimizer, computing update norm (1-state optimizer)"""
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE * N_PER_TH)
+    mask = offsets < n_elements
+
+    g_vals = tl.load(g_ptr + offsets, mask=mask, other=0.0)
+    s1_vals = tl.load(state1_ptr + offsets, mask=mask, other=0.0)
+
+    g_vals = gnorm_scale * g_vals
+
+    if OPTIMIZER_ID == 0:  # MOMENTUM
+        if step == 1:
+            # Cast to fp32 to avoid type mismatch: s1_vals is fp32 but g_vals may be fp16.
+            s1_vals = g_vals.to(tl.float32)
+        else:
+            s1_vals = s1_vals * beta1 + g_vals
+        update_norm = s1_vals * s1_vals
+
+    elif OPTIMIZER_ID == 4:  # LION
+        s1_vals = s1_vals * beta2 + (1.0 - beta2) * g_vals
+        update_norm = s1_vals
+
+    elif OPTIMIZER_ID == 1:  # RMSPROP
+        s1_vals = s1_vals * beta1 + (1.0 - beta1) * g_vals * g_vals
+        update_vals = g_vals / (tl.sqrt(s1_vals) + eps)
+        update_norm = update_vals * update_vals
+
+    elif OPTIMIZER_ID == 2:  # ADAGRAD
+        s1_vals = s1_vals + g_vals * g_vals
+        update_vals = g_vals / (tl.sqrt(s1_vals) + eps)
+        update_norm = update_vals * update_vals
+
+    total_norm = tl.sum(tl.where(mask, update_norm, 0.0))
+
+    tl.atomic_add(unorm_ptr, total_norm)
+
+
+@triton.jit
+def _optimizer_update_2state_32bit_triton_kernel(
+    g_ptr,
+    p_ptr,
+    state1_ptr,
+    state2_ptr,
+    unorm_ptr,
+    max_unorm: tl.constexpr,
+    param_norm,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    beta3,
+    alpha,
+    eps: tl.constexpr,
+    weight_decay: tl.constexpr,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    gnorm_scale: tl.constexpr,
+    skip_zeros,
+    n_elements,
+    OPTIMIZER_ID: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+):
+    """2-state optimizer kernel"""
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE * N_PER_TH)
+    mask = offsets < n_elements
+
+    g_vals = tl.load(g_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+    p_vals = tl.load(p_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+    s1_vals = tl.load(state1_ptr + offsets, mask=mask, other=0.0)
+    s2_vals = tl.load(state2_ptr + offsets, mask=mask, other=0.0)
+
+    if OPTIMIZER_ID == 5:  # ADEMAMIX
+        s3_vals = tl.load(state1_ptr + n_elements + offsets, mask=mask, other=0.0)
+
+    g_vals = gnorm_scale * g_vals
+
+    update_scale = 1.0
+    if max_unorm > 0.0:
+        current_unorm = tl.sqrt(tl.load(unorm_ptr))
+        if current_unorm > max_unorm * param_norm:
+            update_scale = (max_unorm * param_norm) / current_unorm
+
+    if OPTIMIZER_ID == 3:  # ADAM
+        s1_vals = s1_vals * beta1 + (1.0 - beta1) * g_vals
+        s2_vals = s2_vals * beta2 + (1.0 - beta2) * g_vals * g_vals
+
+        correction1 = 1.0 - beta1_step
+        correction2 = tl.sqrt(1.0 - beta2_step)
+        step_size = -lr * correction2 / correction1
+
+        if weight_decay > 0.0:
+            p_vals = p_vals * (1.0 - lr * weight_decay)
+
+        update_val = update_scale * step_size * (s1_vals / (tl.sqrt(s2_vals) + eps * correction2))
+        p_vals = p_vals + update_val
+
+    elif OPTIMIZER_ID == 5:  # ADEMAMIX
+        s1_vals = s1_vals * beta1 + (1.0 - beta1) * g_vals  # m1
+        s3_vals = s3_vals * beta3 + (1.0 - beta3) * g_vals  # m2
+        s2_vals = s2_vals * beta2 + (1.0 - beta2) * g_vals * g_vals  # nu
+
+        correction1 = 1.0 - beta1_step
+        correction2 = tl.sqrt(1.0 - beta2_step)
+
+        if weight_decay > 0.0:
+            p_vals = p_vals * (1.0 - lr * weight_decay)
+
+        mixed_momentum = (s1_vals / correction1) + (alpha * s3_vals)
+        adaptive_term = (tl.sqrt(s2_vals) / correction2) + eps
+        p_vals = p_vals - lr * (mixed_momentum / adaptive_term)
+
+    tl.store(p_ptr + offsets, p_vals, mask=mask)
+    tl.store(state1_ptr + offsets, s1_vals, mask=mask)
+    tl.store(state2_ptr + offsets, s2_vals, mask=mask)
+
+    if OPTIMIZER_ID == 5:  # ADEMAMIX
+        tl.store(state1_ptr + n_elements + offsets, s3_vals, mask=mask)
+
+
+@triton.jit
+def _optimizer_update_1state_32bit_triton_kernel(
+    g_ptr,
+    p_ptr,
+    state1_ptr,
+    state2_ptr,
+    unorm_ptr,
+    max_unorm: tl.constexpr,
+    param_norm,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    beta3,
+    alpha,
+    eps: tl.constexpr,
+    weight_decay: tl.constexpr,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    gnorm_scale: tl.constexpr,
+    skip_zeros,
+    n_elements,
+    OPTIMIZER_ID: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+):
+    """1-state optimizer kernel"""
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE * N_PER_TH)
+    mask = offsets < n_elements
+
+    g_vals = tl.load(g_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+    p_vals = tl.load(p_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+    s1_vals = tl.load(state1_ptr + offsets, mask=mask, other=0.0)
+
+    g_vals = gnorm_scale * g_vals
+    # Coupled (L2) weight decay: fold wd into the gradient. This is correct for
+    # MOMENTUM/RMSPROP/ADAGRAD, but NOT for LION (id 4), which uses *decoupled*
+    # (AdamW-style) weight decay applied to the param directly (see the LION branch
+    # below and Chen et al. 2023). LION is intentionally excluded here.
+    if OPTIMIZER_ID != 4 and weight_decay > 0.0:
+        g_vals = g_vals + p_vals * weight_decay
+
+    update_scale = 1.0
+    if max_unorm > 0.0:
+        current_unorm = tl.sqrt(tl.load(unorm_ptr))
+        if current_unorm > max_unorm * param_norm + eps:
+            update_scale = (max_unorm * param_norm + eps) / current_unorm
+
+    if OPTIMIZER_ID == 0:  # MOMENTUM
+        if step == 1:
+            s1_vals = g_vals
+        else:
+            s1_vals = s1_vals * beta1 + g_vals
+
+        update_val = update_scale * (-lr * s1_vals)
+        p_vals = p_vals + update_val
+
+    elif OPTIMIZER_ID == 4:  # LION
+        # Lion uses decoupled weight decay: shrink the param directly (p *= 1 - lr*wd)
+        # rather than folding wd into the gradient. Matches the 8-bit blockwise kernel,
+        # the default/cpu backends, and the Lion paper (Chen et al. 2023).
+        if weight_decay > 0.0:
+            p_vals = p_vals * (1.0 - lr * weight_decay)
+
+        momentum_update = s1_vals * beta1 + (1.0 - beta1) * g_vals
+        update_val = update_scale * lr * tl.where(momentum_update > 0, 1.0, tl.where(momentum_update < 0, -1.0, 0.0))
+        p_vals = p_vals - update_val
+
+        s1_vals = s1_vals * beta2 + (1.0 - beta2) * g_vals
+
+    elif OPTIMIZER_ID == 1:  # RMSPROP
+        s1_vals = s1_vals * beta1 + (1.0 - beta1) * g_vals * g_vals
+
+        update_val = update_scale * lr * g_vals / (tl.sqrt(s1_vals) + eps)
+        p_vals = p_vals - update_val
+
+    elif OPTIMIZER_ID == 2:  # ADAGRAD
+        s1_vals = s1_vals + g_vals * g_vals
+
+        update_val = lr * g_vals / (tl.sqrt(s1_vals) + eps)
+        p_vals = p_vals - update_val
+
+    tl.store(p_ptr + offsets, p_vals, mask=mask)
+    tl.store(state1_ptr + offsets, s1_vals, mask=mask)
+
+
+name2optimizer_32bit_fn = {
+    "adam": {
+        "preprocess": _optimizer_precondition_2state_32bit,
+        "update": _optimizer_update_2state_32bit_triton_kernel,
+    },
+    "lamb": {
+        "preprocess": _optimizer_precondition_2state_32bit,
+        "update": _optimizer_update_2state_32bit_triton_kernel,
+    },
+    "ademamix": {
+        "preprocess": _optimizer_precondition_2state_32bit,
+        "update": _optimizer_update_2state_32bit_triton_kernel,
+    },
+    "momentum": {
+        "preprocess": _optimizer_precondition_1state_32bit,
+        "update": _optimizer_update_1state_32bit_triton_kernel,
+    },
+    "lars": {
+        "preprocess": _optimizer_precondition_1state_32bit,
+        "update": _optimizer_update_1state_32bit_triton_kernel,
+    },
+    "rmsprop": {
+        "preprocess": _optimizer_precondition_1state_32bit,
+        "update": _optimizer_update_1state_32bit_triton_kernel,
+    },
+    "adagrad": {
+        "preprocess": _optimizer_precondition_1state_32bit,
+        "update": _optimizer_update_1state_32bit_triton_kernel,
+    },
+    "lion": {
+        "preprocess": _optimizer_precondition_1state_32bit,
+        "update": _optimizer_update_1state_32bit_triton_kernel,
+    },
+}
+
+
+def optimizer_update_32bit_impl(
+    optimizer_name: str,
+    g: torch.Tensor,
+    p: torch.Tensor,
+    state1: torch.Tensor,
+    state2: Optional[torch.Tensor],
+    unorm_vec: Optional[torch.Tensor],
+    max_unorm: float,
+    param_norm: float,
+    beta1: float,
+    beta2: float,
+    beta3: float,
+    alpha: float,
+    eps: float,
+    weight_decay: float,
+    step: int,
+    lr: float,
+    gnorm_scale: float = 1.0,
+    skip_zeros=False,
+) -> None:
+    """
+    32-bit optimizer implemented by Triton
+    """
+    if skip_zeros:
+        raise NotImplementedError("skip_zeros is not supported on XPU yet")
+
+    BLOCK_SIZE = 256
+    N_PER_TH = 1  # Number of blocks processed per thread.
+    grid = (triton.cdiv(p.numel(), BLOCK_SIZE * N_PER_TH),)
+    optimizer_id = name2optimizer_id[optimizer_name]
+    fn_preprocess = name2optimizer_32bit_fn[optimizer_name]["preprocess"]
+    fn_update = name2optimizer_32bit_fn[optimizer_name]["update"]
+
+    # In torch=2.7 on XPU there is an issue with libdevice.pow, leading to an error.
+    # For backwards compatibility we precompute the bias correction factors.
+    beta1_step = beta1**step
+    beta2_step = beta2**step
+
+    if optimizer_name == "lion":
+        fn_update[grid](
+            g,
+            p,
+            state1,
+            state2,
+            unorm_vec,
+            max_unorm,
+            param_norm,
+            beta1,
+            beta2,
+            beta3,
+            alpha,
+            eps,
+            weight_decay,
+            step,
+            beta1_step,
+            beta2_step,
+            lr,
+            gnorm_scale,
+            skip_zeros,
+            p.numel(),
+            optimizer_id,
+            BLOCK_SIZE,
+            N_PER_TH,
+            num_warps=2,
+        )
+
+        if max_unorm > 0.0:
+            unorm_vec.zero_()
+            fn_preprocess[grid](
+                g,
+                p,
+                state1,
+                state2,
+                unorm_vec,
+                beta1,
+                beta2,
+                eps,
+                weight_decay,
+                step,
+                beta1_step,
+                beta2_step,
+                lr,
+                gnorm_scale,
+                p.numel(),
+                optimizer_id,
+                BLOCK_SIZE,
+                N_PER_TH,
+                num_warps=2,
+            )
+
+    else:
+        if max_unorm > 0.0:
+            unorm_vec.zero_()
+            fn_preprocess[grid](
+                g,
+                p,
+                state1,
+                state2,
+                unorm_vec,
+                beta1,
+                beta2,
+                eps,
+                weight_decay,
+                step,
+                beta1_step,
+                beta2_step,
+                lr,
+                gnorm_scale,
+                p.numel(),
+                optimizer_id,
+                BLOCK_SIZE,
+                N_PER_TH,
+                num_warps=2,
+            )
+
+        fn_update[grid](
+            g,
+            p,
+            state1,
+            state2,
+            unorm_vec,
+            max_unorm,
+            param_norm,
+            beta1,
+            beta2,
+            beta3,
+            alpha,
+            eps,
+            weight_decay,
+            step,
+            beta1_step,
+            beta2_step,
+            lr,
+            gnorm_scale,
+            skip_zeros,
+            p.numel(),
+            optimizer_id,
+            BLOCK_SIZE,
+            N_PER_TH,
+            num_warps=2,
+        )
+
+
+###########################################
+# Pure torch implementation for reference #
+###########################################
+
+
+@torch.compile
+def _dequantize_blockwise_pytorch(
+    A: torch.Tensor,
+    absmax: torch.Tensor,
+    code: torch.Tensor,
+    blocksize: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    """
+    Pure PyTorch reference implementation for block-wise dequantization.
+    """
+    if A.numel() == 0:
+        return torch.empty_like(A, dtype=dtype)
+
+    A_flat = A.flatten()
+    num_elements = A_flat.numel()
+
+    dequantized_flat = code.to(A.device)[A_flat.long()].to(dtype)
+
+    num_blocks = math.ceil(num_elements / blocksize)
+    pad_len = num_blocks * blocksize - num_elements
+    if pad_len > 0:
+        dequantized_flat = torch.nn.functional.pad(dequantized_flat, (0, pad_len))
+
+    dequantized_blocks = dequantized_flat.reshape(num_blocks, blocksize)
+
+    rescaled_blocks = dequantized_blocks * absmax.unsqueeze(1).to(dtype)
+
+    rescaled_flat = rescaled_blocks.flatten()
+    if pad_len > 0:
+        rescaled_flat = rescaled_flat[:-pad_len]
+
+    return rescaled_flat.reshape(A.shape)
+
+
+@torch.compile
+def _quantize_blockwise_pytorch(
+    A: torch.Tensor,
+    code: torch.Tensor,
+    blocksize: int,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Pure PyTorch reference implementation for block-wise quantization.
+    """
+    if A.numel() == 0:
+        return torch.empty_like(A, dtype=torch.uint8), torch.empty(0, dtype=torch.float32, device=A.device)
+
+    A_flat = A.flatten()
+    num_elements = A_flat.numel()
+
+    num_blocks = math.ceil(num_elements / blocksize)
+
+    pad_len = num_blocks * blocksize - num_elements
+    if pad_len > 0:
+        A_flat = torch.nn.functional.pad(A_flat, (0, pad_len))
+
+    A_blocks = A_flat.reshape(num_blocks, blocksize)
+
+    absmax = torch.max(torch.abs(A_blocks), dim=1, keepdim=True)[0]
+    absmax[absmax == 0] = 1.0
+
+    scaled_blocks = A_blocks / absmax
+
+    # Inefficient but straightforward quantization, takes a lot of memory
+    diff = torch.abs(scaled_blocks.unsqueeze(2) - code.to(A.device))
+    quantized_indices = torch.argmin(diff, dim=2).to(torch.uint8)
+
+    quantized_flat = quantized_indices.flatten()
+    if pad_len > 0:
+        quantized_flat = quantized_flat[:-pad_len]
+
+    return quantized_flat.reshape(A.shape), absmax.flatten()
+
+
+# Main updated function
+def optimizer_update_8bit_blockwise_pytorch(
+    p: torch.Tensor,
+    g: torch.Tensor,
+    state1: torch.Tensor,
+    state2: Optional[torch.Tensor],
+    beta1: float,
+    beta2: float,
+    beta3: float,  # ADEMIX
+    alpha: float,  # ADEMIX
+    eps: float,
+    step: int,
+    lr: float,
+    qmap1: torch.Tensor,
+    qmap2: Optional[torch.Tensor],
+    absmax1: torch.Tensor,
+    absmax2: Optional[torch.Tensor],
+    weight_decay: float,
+    gnorm_scale: float,
+    skip_zeros: bool,
+    # ADEMIX
+    *,
+    optimizer_name: str,
+) -> None:
+    """
+    Pure PyTorch implementation of the 8-bit block-wise optimizer update step.
+    This version ensures high-precision updates for float16 parameters.
+    """
+    if skip_zeros:
+        raise ValueError("skip_zeros is not supported on XPU yet.")
+
+    blocksize = 256
+
+    with torch.no_grad():
+        # Dequantize states to perform updates in 32-bit precision
+        if optimizer_name == "ademamix" and absmax1.ndim == 2:
+            # For AdEMAMix, state1 holds two EMAs, so absmax1 is stacked.
+            s1_1_fp32 = _dequantize_blockwise_pytorch(state1[0], absmax1[0], qmap1, blocksize, torch.float32)
+            s1_2_fp32 = _dequantize_blockwise_pytorch(state1[1], absmax1[1], qmap1, blocksize, torch.float32)
+            state1_fp32 = torch.stack([s1_1_fp32, s1_2_fp32])
+        else:
+            state1_fp32 = _dequantize_blockwise_pytorch(state1, absmax1, qmap1, blocksize, torch.float32)
+
+        state2_fp32 = None
+        if state2 is not None:
+            state2_fp32 = _dequantize_blockwise_pytorch(state2, absmax2, qmap2, blocksize, torch.float32)
+
+        grad = g.float() * gnorm_scale
+
+        # Create a 32-bit copy of the parameter for high-precision updates
+        p_fp32 = p.data.float()
+
+        if optimizer_name == "adam":
+            state1_fp32.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+            state2_fp32.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+
+            bias_correction1 = 1.0 - beta1**step
+            bias_correction2 = 1.0 - beta2**step
+
+            denom = (state2_fp32.sqrt() / math.sqrt(bias_correction2)).add_(eps)
+
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+            p_fp32.addcdiv_(state1_fp32, denom, value=-lr / bias_correction1)
+
+        elif optimizer_name == "ademamix":
+            m1_fp32, m2_fp32 = state1_fp32[0], state1_fp32[1]
+            nu_fp32 = state2_fp32
+
+            m1_fp32.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+            m2_fp32.mul_(beta3).add_(grad, alpha=1.0 - beta3)
+            nu_fp32.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+
+            bias_correction1 = 1.0 - beta1**step
+            bias_correction2 = math.sqrt(1.0 - beta2**step)
+
+            update = (m1_fp32 / bias_correction1 + alpha * m2_fp32) / (nu_fp32.sqrt() / bias_correction2 + eps)
+
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+
+            p_fp32.add_(update, alpha=-lr)
+            state1_fp32 = torch.stack([m1_fp32, m2_fp32])
+
+        elif optimizer_name == "momentum":
+            grad.add_(p_fp32, alpha=weight_decay)
+            if step == 1:
+                state1_fp32.copy_(grad)
+            else:
+                state1_fp32.mul_(beta1).add_(grad)
+            p_fp32.add_(state1_fp32, alpha=-lr)
+
+        elif optimizer_name == "rmsprop":
+            grad.add_(p_fp32, alpha=weight_decay)
+            state1_fp32.mul_(beta1).addcmul_(grad, grad, value=1.0 - beta1)
+            p_fp32.addcdiv_(grad, state1_fp32.sqrt().add_(eps), value=-lr)
+
+        elif optimizer_name == "lion":
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+
+            update_dir = torch.sign(state1_fp32.mul(beta1) + grad.mul(1.0 - beta1))
+            p_fp32.add_(update_dir, alpha=-lr)
+
+            state1_fp32.mul_(beta2).add_(grad, alpha=1.0 - beta2)
+
+        elif optimizer_name == "adagrad":
+            grad.add_(p_fp32, alpha=weight_decay)
+            state1_fp32.addcmul_(grad, grad, value=1.0)
+            p_fp32.addcdiv_(grad, state1_fp32.sqrt().add_(eps), value=-lr)
+
+        else:
+            raise NotImplementedError(
+                f"Pure PyTorch implementation for optimizer '{optimizer_name}' is not available."
+            )
+
+        # Copy the updated 32-bit parameter back to the original tensor
+        p.data.copy_(p_fp32)
+
+        # Re-quantize states and update state tensors in-place
+        if optimizer_name == "ademamix":
+            new_m1_8bit, new_absmax_m1 = _quantize_blockwise_pytorch(state1_fp32[0], qmap1, blocksize)
+            new_m2_8bit, new_absmax_m2 = _quantize_blockwise_pytorch(state1_fp32[1], qmap1, blocksize)
+            state1[0].copy_(new_m1_8bit)
+            state1[1].copy_(new_m2_8bit)
+            absmax1[0].copy_(new_absmax_m1)
+            absmax1[1].copy_(new_absmax_m2)
+
+            new_state2_8bit, new_absmax2 = _quantize_blockwise_pytorch(state2_fp32, qmap2, blocksize)
+            state2.copy_(new_state2_8bit)
+            absmax2.copy_(new_absmax2)
+        else:
+            new_state1_8bit, new_absmax1 = _quantize_blockwise_pytorch(state1_fp32, qmap1, blocksize)
+            state1.copy_(new_state1_8bit)
+            absmax1.copy_(new_absmax1)
+
+            if state2_fp32 is not None:
+                new_state2_8bit, new_absmax2 = _quantize_blockwise_pytorch(state2_fp32, qmap2, blocksize)
+                state2.copy_(new_state2_8bit)
+                absmax2.copy_(new_absmax2)
+
+
+#######################################
+# Mixed torch + triton implementation #
+#######################################
+
+
+# Much more memory efficient due to using triton for quantization/dequantization
+def optimizer_update_8bit_blockwise_triton_quant(
+    p: torch.Tensor,
+    g: torch.Tensor,
+    state1: torch.Tensor,
+    state2: Optional[torch.Tensor],
+    beta1: float,
+    beta2: float,
+    beta3: float,  # ADEMIX
+    alpha: float,  # ADEMIX
+    eps: float,
+    step: int,
+    lr: float,
+    qmap1: torch.Tensor,
+    qmap2: Optional[torch.Tensor],
+    absmax1: torch.Tensor,
+    absmax2: Optional[torch.Tensor],
+    weight_decay: float,
+    gnorm_scale: float,
+    skip_zeros: bool,
+    # ADEMIX
+    *,
+    optimizer_name: str,
+) -> None:
+    """
+    Pure PyTorch implementation of the 8-bit block-wise optimizer update step.
+    This version ensures high-precision updates for float16 parameters.
+    """
+    if skip_zeros and not torch.any(g):
+        return
+
+    blocksize = 256
+    grad = g.float() * gnorm_scale
+
+    with torch.no_grad():
+        # Create a 32-bit copy of the parameter for high-precision updates
+        p_fp32 = p.data.float()
+
+        # Dequantize states to perform updates in 32-bit precision
+        if optimizer_name == "ademamix" and absmax1.ndim == 2:
+            # For AdEMAMix, state1 holds two EMAs, so absmax1 is stacked.
+            s1_1_fp32 = dequant_8bit_blockwise(state1[0], absmax1[0], qmap1, blocksize, dtype=torch.float32)
+            s1_2_fp32 = dequant_8bit_blockwise(state1[1], absmax1[1], qmap1, blocksize, dtype=torch.float32)
+            state1_fp32 = torch.stack([s1_1_fp32, s1_2_fp32])
+        else:
+            state1_fp32 = dequant_8bit_blockwise(state1, absmax1, qmap1, blocksize, dtype=torch.float32)
+
+        state2_fp32 = None
+        if state2 is not None:
+            state2_fp32 = dequant_8bit_blockwise(state2, absmax2, qmap2, blocksize, dtype=torch.float32)
+
+        # Apply optimizer-specific update logic
+        if optimizer_name == "adam":
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+
+            state1_fp32.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+            state2_fp32.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+
+            bias_correction1 = 1.0 - beta1**step
+            bias_correction2 = 1.0 - beta2**step
+
+            denom = (state2_fp32.sqrt() / math.sqrt(bias_correction2)).add_(eps)
+            p_fp32.addcdiv_(state1_fp32, denom, value=-lr / bias_correction1)
+
+        elif optimizer_name == "ademamix":
+            m1_fp32, m2_fp32 = state1_fp32[0], state1_fp32[1]
+            nu_fp32 = state2_fp32
+
+            m1_fp32.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+            m2_fp32.mul_(beta3).add_(grad, alpha=1.0 - beta3)
+            nu_fp32.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+
+            bias_correction1 = 1.0 - beta1**step
+            bias_correction2 = math.sqrt(1.0 - beta2**step)
+
+            update = (m1_fp32 / bias_correction1 + alpha * m2_fp32) / (nu_fp32.sqrt() / bias_correction2 + eps)
+
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+
+            p_fp32.add_(update, alpha=-lr)
+            state1_fp32 = torch.stack([m1_fp32, m2_fp32])
+
+        elif optimizer_name == "momentum":
+            grad.add_(p_fp32, alpha=weight_decay)
+            if step == 1:
+                state1_fp32.copy_(grad)
+            else:
+                state1_fp32.mul_(beta1).add_(grad)
+            p_fp32.add_(state1_fp32, alpha=-lr)
+
+        elif optimizer_name == "rmsprop":
+            grad.add_(p_fp32, alpha=weight_decay)
+            state1_fp32.mul_(beta1).addcmul_(grad, grad, value=1.0 - beta1)
+            p_fp32.addcdiv_(grad, state1_fp32.sqrt().add_(eps), value=-lr)
+
+        elif optimizer_name == "lion":
+            if weight_decay > 0.0:
+                p_fp32.mul_(1.0 - lr * weight_decay)
+
+            update_dir = torch.sign(state1_fp32.mul(beta1) + grad.mul(1.0 - beta1))
+            p_fp32.add_(update_dir, alpha=-lr)
+
+            state1_fp32.mul_(beta2).add_(grad, alpha=1.0 - beta2)
+
+        elif optimizer_name == "adagrad":
+            grad.add_(p_fp32, alpha=weight_decay)
+            state1_fp32.addcmul_(grad, grad, value=1.0)
+            p_fp32.addcdiv_(grad, state1_fp32.sqrt().add_(eps), value=-lr)
+
+        else:
+            raise NotImplementedError(
+                f"Pure PyTorch implementation for optimizer '{optimizer_name}' is not available."
+            )
+
+        # Copy the updated 32-bit parameter back to the original tensor
+        p.data.copy_(p_fp32)
+
+        # Re-quantize states and update state tensors in-place
+        if optimizer_name == "ademamix":
+            new_m1_8bit, new_absmax_m1 = quantize_blockwise_triton(state1_fp32[0], qmap1, blocksize)
+            new_m2_8bit, new_absmax_m2 = quantize_blockwise_triton(state1_fp32[1], qmap1, blocksize)
+            state1[0].copy_(new_m1_8bit)
+            state1[1].copy_(new_m2_8bit)
+            absmax1[0].copy_(new_absmax_m1)
+            absmax1[1].copy_(new_absmax_m2)
+
+            new_state2_8bit, new_absmax2 = quantize_blockwise_triton(state2_fp32, qmap2, blocksize)
+            state2.copy_(new_state2_8bit)
+            absmax2.copy_(new_absmax2)
+        else:
+            new_state1_8bit, new_absmax1 = quantize_blockwise_triton(state1_fp32, qmap1, blocksize)
+            state1.copy_(new_state1_8bit)
+            absmax1.copy_(new_absmax1)
+
+            if state2_fp32 is not None:
+                new_state2_8bit, new_absmax2 = quantize_blockwise_triton(state2_fp32, qmap2, blocksize)
+                state2.copy_(new_state2_8bit)
+                absmax2.copy_(new_absmax2)
+
+
+#########################
+# Triton implementation #
+#########################
+
+
+@triton.jit
+def _optimizer_update_1state_8bit_blockwise_triton_kernel(
+    # Tensors
+    p_ptr,
+    g_ptr,
+    state1_ptr,
+    state2_ptr,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    beta3,
+    alpha,
+    eps: tl.constexpr,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    qmap1_ptr,
+    qmap2_ptr,
+    absmax1_ptr,
+    absmax2_ptr,
+    weight_decay,
+    gnorm_scale,
+    # Meta-parameters
+    n_elements,
+    BLOCK_SIZE_N: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+    OPTIMIZER_ID: tl.constexpr,
+):
+    """
+    Triton kernel for 8-bit optimizers that use one momentum state.
+    Supports: Momentum, RMSprop, Adagrad, Lion.
+    """
+    # 1. Boilerplate: pid, offsets, mask
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N * N_PER_TH)
+    mask = offsets < n_elements
+
+    # 2. Load and dequantize tensors
+    g = tl.load(g_ptr + offsets, mask=mask, other=0.0).to(tl.float32) * gnorm_scale
+    p = tl.load(p_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+    s1 = dequant_8bit_blockwise_kernel_util(state1_ptr, offsets, qmap1_ptr, absmax1_ptr, mask, BLOCK_SIZE_N)
+
+    # 3. Optimizer-specific updates
+    # LION
+    if weight_decay > 0.0 and OPTIMIZER_ID == 2:
+        p *= 1.0 - lr * weight_decay
+    # Apply weight decay for momentum, rmsprop, adagrad
+    elif weight_decay > 0.0:
+        g += p * weight_decay
+
+    # Momentum update
+    if OPTIMIZER_ID == 0:  # MOMENTUM
+        if step == 1:
+            s1 = g
+        else:
+            s1 = s1 * beta1 + g
+        p -= lr * s1
+
+    # RMSprop update
+    elif OPTIMIZER_ID == 1:  # RMSPROP
+        s1 = s1 * beta1 + (1.0 - beta1) * g * g
+        p -= lr * (g / (tl.sqrt(s1) + eps))
+
+    # Adagrad update
+    elif OPTIMIZER_ID == 2:  # ADAGRAD
+        s1 += g * g
+        p -= lr * (g / (tl.sqrt(s1) + eps))
+
+    # Lion update
+    elif OPTIMIZER_ID == 4:  # LION
+        val = s1 * beta1 + (1.0 - beta1) * g
+        update = tl.where(val > 0.0, 1.0, tl.where(val < 0.0, -1.0, 0.0))
+        p -= lr * update
+        s1 = s1 * beta2 + (1.0 - beta2) * g
+
+    # 4. Store updated parameter and requantized state
+    tl.store(p_ptr + offsets, p.to(p_ptr.dtype.element_ty), mask=mask)
+    s1_codes, new_absmax1 = quantize_8bit_blockwise_kernel_util(s1, qmap1_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+    tl.store(state1_ptr + offsets, s1_codes, mask=mask)
+    tl.store(absmax1_ptr + block_start_idx + tl.arange(0, N_PER_TH), new_absmax1)
+
+
+@triton.jit
+def _optimizer_update_2state_8bit_blockwise_triton_kernel(
+    # Tensors
+    p_ptr,
+    g_ptr,
+    state1_ptr,
+    state2_ptr,
+    beta1: tl.constexpr,
+    beta2: tl.constexpr,
+    # ademamix changes alpha and beta3
+    beta3,
+    # ademamix changes alpha and beta3
+    alpha,
+    eps: tl.constexpr,
+    step,
+    beta1_step,
+    beta2_step,
+    lr,
+    qmap1_ptr,
+    qmap2_ptr,
+    absmax1_ptr,
+    absmax2_ptr,
+    weight_decay: tl.constexpr,
+    gnorm_scale: tl.constexpr,
+    # Meta-parameters
+    n_elements,
+    BLOCK_SIZE_N: tl.constexpr,
+    N_PER_TH: tl.constexpr,
+    OPTIMIZER_ID: tl.constexpr,
+):
+    """
+    Triton kernel for 8-bit optimizers that use two momentum states.
+    Supports: Adam, AdEMAMix.
+    """
+    # 1. Boilerplate: pid, offsets, mask
+    pid = tl.program_id(axis=0)
+    block_start_idx = pid * N_PER_TH
+    offsets = block_start_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N * N_PER_TH)
+    mask = offsets < n_elements
+
+    # 2. Load and dequantize tensors
+    g = tl.load(g_ptr + offsets, mask=mask, other=0.0).to(tl.float32) * gnorm_scale
+    p = tl.load(p_ptr + offsets, mask=mask, other=0.0).to(tl.float32)
+
+    # 3. Optimizer-specific updates
+    if OPTIMIZER_ID == 3:  # ADAM
+        s1 = dequant_8bit_blockwise_kernel_util(state1_ptr, offsets, qmap1_ptr, absmax1_ptr, mask, BLOCK_SIZE_N)
+        s2 = dequant_8bit_blockwise_kernel_util(state2_ptr, offsets, qmap2_ptr, absmax2_ptr, mask, BLOCK_SIZE_N)
+
+        s1 = s1 * beta1 + (1.0 - beta1) * g
+        s2 = s2 * beta2 + (1.0 - beta2) * g * g
+
+        # In torch=2.7 on XPU there is an issue with libdevice.pow, leading to an error.
+        # For backwards compatibility we precompute the bias correction factors.
+        # bias_correction1 = 1.0 - libdevice.pow(beta1, step)
+        # bias_correction2 = 1.0 - libdevice.pow(beta2, step)
+        bias_correction1 = 1.0 - beta1_step
+        bias_correction2 = 1.0 - beta2_step
+
+        if weight_decay > 0.0:
+            p *= 1.0 - lr * weight_decay
+
+        denom = tl.sqrt(s2) / tl.sqrt(bias_correction2) + eps
+        p -= (lr / bias_correction1) * (s1 / denom)
+
+        # Store updated parameter
+        tl.store(p_ptr + offsets, p.to(p_ptr.dtype.element_ty), mask=mask)
+
+        # Requantize and store states
+        s1_codes, new_absmax1 = quantize_8bit_blockwise_kernel_util(s1, qmap1_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+        tl.store(state1_ptr + offsets, s1_codes, mask=mask)
+        tl.store(absmax1_ptr + block_start_idx + tl.arange(0, N_PER_TH), new_absmax1)
+
+        s2_codes, new_absmax2 = quantize_8bit_blockwise_kernel_util(s2, qmap2_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+        tl.store(state2_ptr + offsets, s2_codes, mask=mask)
+        tl.store(absmax2_ptr + block_start_idx + tl.arange(0, N_PER_TH), new_absmax2)
+
+    elif OPTIMIZER_ID == 5:  # ADEMAMIX
+        # AdEMAMix has a stacked state1 (m1, m2) and state2 (nu)
+        m1 = dequant_8bit_blockwise_kernel_util(state1_ptr, offsets, qmap1_ptr, absmax1_ptr, mask, BLOCK_SIZE_N)
+        m2 = dequant_8bit_blockwise_kernel_util(
+            state1_ptr + n_elements,
+            offsets,
+            qmap1_ptr,
+            absmax1_ptr + n_elements // BLOCK_SIZE_N,
+            mask,
+            BLOCK_SIZE_N,
+        )
+        nu = dequant_8bit_blockwise_kernel_util(state2_ptr, offsets, qmap2_ptr, absmax2_ptr, mask, BLOCK_SIZE_N)
+
+        m1 = m1 * beta1 + (1.0 - beta1) * g
+        m2 = m2 * beta3 + (1.0 - beta3) * g
+        nu = nu * beta2 + (1.0 - beta2) * g * g
+
+        # In torch=2.7 on XPU there is an issue with libdevice.pow, leading to an error.
+        # For backwards compatibility we precompute the bias correction factors.
+        # bias_correction1 = 1.0 - libdevice.pow(beta1, step)
+        # bias_correction2 = tl.sqrt(1.0 - libdevice.pow(beta2, step))
+        bias_correction1 = 1.0 - beta1_step
+        bias_correction2 = tl.sqrt(1.0 - beta2_step)
+
+        update = (m1 / bias_correction1 + alpha * m2) / (tl.sqrt(nu) / bias_correction2 + eps)
+
+        if weight_decay > 0.0:
+            p *= 1.0 - lr * weight_decay
+
+        p -= lr * update
+
+        # Store updated parameter
+        tl.store(p_ptr + offsets, p.to(p_ptr.dtype.element_ty), mask=mask)
+
+        # Requantize and store all three states
+        m1_codes, new_absmax_m1 = quantize_8bit_blockwise_kernel_util(m1, qmap1_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+        tl.store(state1_ptr + offsets, m1_codes, mask=mask)
+        tl.store(absmax1_ptr + block_start_idx + tl.arange(0, N_PER_TH), new_absmax_m1)
+
+        m2_codes, new_absmax_m2 = quantize_8bit_blockwise_kernel_util(m2, qmap1_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+        tl.store(state1_ptr + n_elements + offsets, m2_codes, mask=mask)
+        tl.store(
+            absmax1_ptr + block_start_idx + tl.arange(0, N_PER_TH) + n_elements // BLOCK_SIZE_N,
+            new_absmax_m2,
+        )
+
+        nu_codes, new_absmax_nu = quantize_8bit_blockwise_kernel_util(nu, qmap2_ptr, 256, BLOCK_SIZE_N, N_PER_TH)
+        tl.store(state2_ptr + offsets, nu_codes, mask=mask)
+        tl.store(absmax2_ptr + block_start_idx + tl.arange(0, N_PER_TH), new_absmax_nu)
+
+
+name2optimizer_fn = {
+    "momentum": _optimizer_update_1state_8bit_blockwise_triton_kernel,
+    "lars": _optimizer_update_1state_8bit_blockwise_triton_kernel,
+    "rmsprop": _optimizer_update_1state_8bit_blockwise_triton_kernel,
+    "adagrad": _optimizer_update_1state_8bit_blockwise_triton_kernel,
+    "adam": _optimizer_update_2state_8bit_blockwise_triton_kernel,
+    "lamb": _optimizer_update_2state_8bit_blockwise_triton_kernel,
+    "lion": _optimizer_update_1state_8bit_blockwise_triton_kernel,
+    "ademamix": _optimizer_update_2state_8bit_blockwise_triton_kernel,
+}
+
+
+def optimizer_update_8bit_blockwise_impl(
+    optimizer_name: str,
+    g: torch.Tensor,
+    p: torch.Tensor,
+    state1: torch.Tensor,
+    state2: Optional[torch.Tensor],
+    beta1: float,
+    beta2: float,
+    beta3: float,
+    alpha: float,
+    eps: float,
+    step: int,
+    lr: float,
+    qmap1: torch.Tensor,
+    qmap2: Optional[torch.Tensor],
+    absmax1: torch.Tensor,
+    absmax2: Optional[torch.Tensor],
+    weight_decay: float = 0.0,
+    gnorm_scale: float = 1.0,
+    skip_zeros=False,
+) -> None:
+    if skip_zeros:
+        raise NotImplementedError("skip_zeros is not supported on XPU yet")
+
+    if optimizer_name == "ademamix":
+        # Handle AdEMAMIX's stacked state tensors
+        if state1.dim() < 2 or state1.shape[0] != 2:
+            raise ValueError(
+                f"For ademamix, state1 must be a stacked tensor of shape (2, ...), but got {state1.shape}"
+            )
+        if absmax1.dim() < 2 or absmax1.shape[0] != 2:
+            raise ValueError(
+                f"For ademamix, absmax1 must be a stacked tensor of shape (2, ...), but got {absmax1.shape}"
+            )
+
+    BLOCK_SIZE = 256
+    N_PER_TH = 1  # Number of blocks processed per thread.
+    grid = (triton.cdiv(p.numel(), BLOCK_SIZE * N_PER_TH),)
+    fn = name2optimizer_fn[optimizer_name]
+    optimizer_id = name2optimizer_id[optimizer_name]
+
+    # In torch=2.7 on XPU there is an issue with libdevice.pow, leading to an error.
+    # For backwards compatibility we precompute the bias correction factors.
+    beta1_step = beta1**step
+    beta2_step = beta2**step
+
+    fn[grid](
+        p,
+        g,
+        state1,
+        state2,
+        beta1,
+        beta2,
+        beta3,
+        alpha,
+        eps,
+        step,
+        beta1_step,
+        beta2_step,
+        lr,
+        qmap1,
+        qmap2,
+        absmax1,
+        absmax2,
+        weight_decay,
+        gnorm_scale,
+        p.numel(),
+        BLOCK_SIZE_N=BLOCK_SIZE,
+        N_PER_TH=N_PER_TH,
+        OPTIMIZER_ID=optimizer_id,
+        num_warps=2,
+    )
+
+
+# optimizer_update_8bit_blockwise_impl = optimizer_update_8bit_blockwise_pytorch
+# optimizer_update_8bit_blockwise_impl = torch.compile(optimizer_update_8bit_blockwise_pytorch_impl)
+# optimizer_update_8bit_blockwise_impl = optimizer_update_8bit_blockwise_triton_quant
+# optimizer_update_8bit_blockwise_impl = torch.compile(optimizer_update_8bit_blockwise_triton_quant)
+optimizer_update_8bit_blockwise_impl = optimizer_update_8bit_blockwise_impl