Deepen offload prefetch pipeline with per-slot events

Replace the 2-slot ping-pong + single global "compute-started" event
with a depth-N ring buffer where each transfer waits only on the slot
it's reusing (D layers back) instead of the most-recent compute. Applies
to forward and backward, Linear and Conv. Depth is tunable via
AI_TOOLKIT_OFFLOAD_DEPTH (default 4).

Bit-exact vs non-offload (output, grad_input, weight grads). No speedup
on a bandwidth-bound PCIe link (already saturated at depth 2), but the
cleaner per-slot design removes the fragile shared-event serialization
and lets deeper prefetch help on faster buses.
This commit is contained in:
Jaret Burkett
2026-06-07 16:07:13 -06:00
parent cdff6e36aa
commit c78b1404e3
2 changed files with 486 additions and 149 deletions

View File

@@ -6,6 +6,8 @@ https://github.com/lodestone-rock/RamTorch/blob/main/ramtorch/modules/linear.py
I simply modified it to work with a memory management model and with AI Toolkit's models
"""
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
@@ -18,6 +20,12 @@ if TYPE_CHECKING:
# --- Per-device global state registry ---
_DEVICE_STATE = {}
# How many layers deep to prefetch weights. The old ping-pong used 2 slots, which
# only lets one transfer overlap one compute (1-deep). A deeper ring lets Python
# enqueue several layers ahead so the H2D stream stays saturated instead of
# stalling on a per-layer sync. Override with AI_TOOLKIT_OFFLOAD_DEPTH.
PIPELINE_DEPTH = int(os.environ.get("AI_TOOLKIT_OFFLOAD_DEPTH", "4"))
def _get_device_state(device: torch.device):
"""Get or initialize per-device state."""
@@ -31,31 +39,101 @@ def _get_device_state(device: torch.device):
return _DEVICE_STATE[device]
if device not in _DEVICE_STATE:
d = max(2, PIPELINE_DEPTH)
with torch.cuda.device(device):
_DEVICE_STATE[device] = {
# streams & events
"depth": d,
# streams
"transfer_stream": torch.cuda.Stream(device=device),
"transfer_grad_stream": torch.cuda.Stream(device=device),
"transfer_forward_finished_event": torch.cuda.Event(),
"compute_forward_start_event": torch.cuda.Event(),
"transfer_backward_finished_event": torch.cuda.Event(),
"transfer_weight_backward_finished_event": torch.cuda.Event(),
"compute_backward_start_event": torch.cuda.Event(),
"compute_backward_finished_event": torch.cuda.Event(),
# ping-pong buffers
"w_buffers": [None, None],
"b_buffers": [None, None],
"w_bwd_buffers": [None, None],
# device-side staging for grads to be sent to CPU
"w_grad_buffers": [None, None],
"b_grad_buffers": [None, None],
# clocks
# forward weight ring: slot_ready = H2D done, slot_free = compute
# that consumed the slot done (so it can be overwritten).
"w_buffers": [None] * d,
"b_buffers": [None] * d,
"fwd_slot_ready": [torch.cuda.Event() for _ in range(d)],
"fwd_slot_free": [torch.cuda.Event() for _ in range(d)],
"forward_clk": 0,
# backward weight ring (re-fetch for grad-input).
"w_bwd_buffers": [None] * d,
"bwd_slot_ready": [torch.cuda.Event() for _ in range(d)],
"bwd_slot_free": [torch.cuda.Event() for _ in range(d)],
"backward_clk": 0,
# backward grad-staging ring (device-side grads -> CPU).
"w_grad_buffers": [None] * d,
"b_grad_buffers": [None] * d,
"grad_compute_done": [torch.cuda.Event() for _ in range(d)],
"grad_xfer_done": [torch.cuda.Event() for _ in range(d)],
}
return _DEVICE_STATE[device]
# ---- ring-buffer staging helpers -----------------------------------------
#
# Each transfer waits only on the event for the *specific slot* it is about to
# overwrite (the compute that used that slot D layers ago), not on a single
# global "compute started" event. With D slots that prior compute is long done,
# so the transfer stream never actually stalls and stays D layers ahead of
# compute. This is the deeper-pipeline + relaxed-dependency change in one.
def _stage_forward_weight(state, device, materialize, weight_cpu, bias_cpu):
"""H2D the next forward weight (+bias) into its ring slot; return (idx, w, b).
Caller runs compute, then calls _release_forward_slot(state, idx)."""
d = state["depth"]
idx = state["forward_clk"]
state["forward_clk"] = (idx + 1) % d
ts = state["transfer_stream"]
with torch.cuda.stream(ts):
ts.wait_event(state["fwd_slot_free"][idx])
state["w_buffers"][idx] = materialize(weight_cpu, device)
state["b_buffers"][idx] = (
bias_cpu.to(device, non_blocking=True) if bias_cpu is not None else None
)
state["fwd_slot_ready"][idx].record()
torch.cuda.current_stream().wait_event(state["fwd_slot_ready"][idx])
return idx, state["w_buffers"][idx], state["b_buffers"][idx]
def _release_forward_slot(state, idx):
# Slot is reusable once the compute stream finishes the op that read it.
state["fwd_slot_free"][idx].record()
def _stage_backward_weight(state, device, materialize, weight_cpu):
"""H2D the next backward weight into its ring slot; return (idx, w).
Caller runs grad-input compute, then _release_backward_weight_slot."""
d = state["depth"]
idx = state["backward_clk"]
state["backward_clk"] = (idx + 1) % d
ts = state["transfer_stream"]
with torch.cuda.stream(ts):
ts.wait_event(state["bwd_slot_free"][idx])
state["w_bwd_buffers"][idx] = materialize(weight_cpu)
state["bwd_slot_ready"][idx].record()
torch.cuda.current_stream().wait_event(state["bwd_slot_ready"][idx])
return idx, state["w_bwd_buffers"][idx]
def _release_backward_weight_slot(state, idx):
state["bwd_slot_free"][idx].record()
def _stage_grads_to_cpu(state, idx, grad_w_gpu, grad_b_gpu):
"""Copy freshly-computed device grads (in staging slot idx) to CPU on the
grad stream, overlapping the next H2D. Returns (grad_w_cpu, grad_b_cpu)."""
gs = state["transfer_grad_stream"]
state["grad_compute_done"][idx].record() # on the compute stream
grad_w_cpu = grad_b_cpu = None
with torch.cuda.stream(gs):
gs.wait_event(state["grad_compute_done"][idx])
if grad_w_gpu is not None:
grad_w_cpu = grad_w_gpu.to("cpu", non_blocking=True)
if grad_b_gpu is not None:
grad_b_cpu = grad_b_gpu.to("cpu", non_blocking=True)
state["grad_xfer_done"][idx].record()
return grad_w_cpu, grad_b_cpu
# (ADD) detect torchao wrapper tensors
def _is_ao_quantized_tensor(t: Optional[torch.Tensor]) -> bool:
if t is None:
@@ -209,24 +287,11 @@ class _BouncingLinearFn(torch.autograd.Function):
return out.to(x.device)
state = _get_device_state(device)
ts = state["transfer_stream"]
w_bufs, b_bufs = state["w_buffers"], state["b_buffers"]
ev_tx_f = state["transfer_forward_finished_event"]
ev_cu_s = state["compute_forward_start_event"]
idx = state["forward_clk"]
with torch.cuda.stream(ts):
ts.wait_event(ev_cu_s)
w_bufs[idx] = _materialize_linear_weight(weight_cpu, device)
b_bufs[idx] = (
bias_cpu.to(device, non_blocking=True) if bias_cpu is not None else None
)
state["forward_clk"] ^= 1
ev_tx_f.record()
torch.cuda.current_stream().wait_event(ev_tx_f)
ev_cu_s.record()
out = F.linear(x, w_bufs[idx], b_bufs[idx])
idx, w_gpu, b_gpu = _stage_forward_weight(
state, device, _materialize_linear_weight, weight_cpu, bias_cpu
)
out = F.linear(x, w_gpu, b_gpu)
_release_forward_slot(state, idx)
ctx.save_for_backward(x, weight_cpu, bias_cpu)
ctx.device = device
@@ -268,19 +333,6 @@ class _BouncingLinearFn(torch.autograd.Function):
return grad_input.to(grad_out.device), grad_weight, grad_bias, None
state = _get_device_state(device)
transfer_stream = state["transfer_stream"]
transfer_grad_stream = state["transfer_grad_stream"]
w_bwd_buffers = state["w_bwd_buffers"]
w_grad_buffers = state["w_grad_buffers"]
b_grad_buffers = state["b_grad_buffers"]
ev_tx_b = state["transfer_backward_finished_event"]
ev_tx_w_bwd_done = state["transfer_weight_backward_finished_event"]
ev_cu_b_start = state["compute_backward_start_event"]
ev_cu_b_finish = state["compute_backward_finished_event"]
idx = state["backward_clk"]
# GPU-side dequant/cast for quantized; float path unchanged
def _materialize_for_bwd(cpu_w):
@@ -297,45 +349,35 @@ class _BouncingLinearFn(torch.autograd.Function):
w = cpu_w.to(device, non_blocking=True)
return w
with torch.cuda.stream(transfer_stream):
transfer_stream.wait_event(ev_cu_b_start)
w_bwd_buffers[idx] = _materialize_for_bwd(weight_cpu)
state["backward_clk"] ^= 1
ev_tx_b.record()
torch.cuda.current_stream().wait_event(ev_tx_b)
ev_cu_b_start.record()
idx, w_bwd = _stage_backward_weight(
state, device, _materialize_for_bwd, weight_cpu
)
# grad wrt input (GPU)
grad_input = grad_out.to(dtype=target_dtype) @ w_bwd_buffers[idx]
grad_input = grad_out.to(dtype=target_dtype) @ w_bwd
_release_backward_weight_slot(state, idx)
# ensure previous grad-to-CPU transfer that used this slot finished
torch.cuda.current_stream().wait_event(ev_tx_w_bwd_done)
# compute grads if float masters exist
# compute grads if float masters exist (frozen/quantized bases skip this)
grad_weight = None
grad_bias = None
if (
need_w = (
getattr(weight_cpu, "requires_grad", False)
and weight_cpu.dtype.is_floating_point
):
w_grad_buffers[idx] = grad_out.flatten(0, -2).T @ x.flatten(0, -2)
if bias_cpu is not None and getattr(bias_cpu, "requires_grad", False):
reduce_dims = tuple(range(grad_out.ndim - 1))
b_grad_buffers[idx] = grad_out.sum(dim=reduce_dims)
ev_cu_b_finish.record()
with torch.cuda.stream(transfer_grad_stream):
transfer_grad_stream.wait_event(ev_cu_b_finish)
if (
getattr(weight_cpu, "requires_grad", False)
and weight_cpu.dtype.is_floating_point
):
grad_weight = w_grad_buffers[idx].to("cpu", non_blocking=True)
if bias_cpu is not None and getattr(bias_cpu, "requires_grad", False):
grad_bias = b_grad_buffers[idx].to("cpu", non_blocking=True)
state["transfer_weight_backward_finished_event"].record()
)
need_b = bias_cpu is not None and getattr(bias_cpu, "requires_grad", False)
if need_w or need_b:
# ensure the prior grad D2H using this staging slot finished
torch.cuda.current_stream().wait_event(state["grad_xfer_done"][idx])
w_grad_gpu = b_grad_gpu = None
if need_w:
w_grad_gpu = grad_out.flatten(0, -2).T @ x.flatten(0, -2)
state["w_grad_buffers"][idx] = w_grad_gpu
if need_b:
b_grad_gpu = grad_out.sum(dim=tuple(range(grad_out.ndim - 1)))
state["b_grad_buffers"][idx] = b_grad_gpu
grad_weight, grad_bias = _stage_grads_to_cpu(
state, idx, w_grad_gpu, b_grad_gpu
)
return grad_input.to(dtype=grad_out.dtype), grad_weight, grad_bias, None
@@ -389,24 +431,11 @@ class _BouncingConv2dFn(torch.autograd.Function):
return out.to(x.device)
state = _get_device_state(device)
ts = state["transfer_stream"]
w_bufs, b_bufs = state["w_buffers"], state["b_buffers"]
ev_tx_f = state["transfer_forward_finished_event"]
ev_cu_s = state["compute_forward_start_event"]
idx = state["forward_clk"]
with torch.cuda.stream(ts):
ts.wait_event(ev_cu_s)
w_bufs[idx] = _materialize_conv_weight(weight_cpu, device)
b_bufs[idx] = (
bias_cpu.to(device, non_blocking=True) if bias_cpu is not None else None
)
state["forward_clk"] ^= 1
ev_tx_f.record()
torch.cuda.current_stream().wait_event(ev_tx_f)
ev_cu_s.record()
out = F.conv2d(x, w_bufs[idx], b_bufs[idx], stride, padding, dilation, groups)
idx, w_gpu, b_gpu = _stage_forward_weight(
state, device, _materialize_conv_weight, weight_cpu, bias_cpu
)
out = F.conv2d(x, w_gpu, b_gpu, stride, padding, dilation, groups)
_release_forward_slot(state, idx)
ctx.save_for_backward(x, weight_cpu, bias_cpu)
ctx.meta = (device, stride, padding, dilation, groups, target_dtype)
@@ -475,19 +504,6 @@ class _BouncingConv2dFn(torch.autograd.Function):
)
state = _get_device_state(device)
transfer_stream = state["transfer_stream"]
transfer_grad_stream = state["transfer_grad_stream"]
w_bwd_buffers = state["w_bwd_buffers"]
w_grad_buffers = state["w_grad_buffers"]
b_grad_buffers = state["b_grad_buffers"]
ev_tx_b = state["transfer_backward_finished_event"]
ev_tx_w_bwd_done = state["transfer_weight_backward_finished_event"]
ev_cu_b_start = state["compute_backward_start_event"]
ev_cu_b_finish = state["compute_backward_finished_event"]
idx = state["backward_clk"]
# GPU-side dequant/cast for quantized; float path unchanged
def _materialize_for_bwd(cpu_w):
@@ -504,63 +520,51 @@ class _BouncingConv2dFn(torch.autograd.Function):
w = cpu_w.to(device, non_blocking=True)
return w
# Stage weights for input-grad compute
with torch.cuda.stream(transfer_stream):
transfer_stream.wait_event(ev_cu_b_start)
w_bwd_buffers[idx] = _materialize_for_bwd(weight_cpu)
state["backward_clk"] ^= 1
ev_tx_b.record()
torch.cuda.current_stream().wait_event(ev_tx_b)
ev_cu_b_start.record()
idx, w_bwd = _stage_backward_weight(
state, device, _materialize_for_bwd, weight_cpu
)
from torch.nn.grad import conv2d_input, conv2d_weight # type: ignore
grad_input = conv2d_input(
x.shape,
w_bwd_buffers[idx],
w_bwd,
grad_out.to(dtype=target_dtype),
stride=stride,
padding=padding,
dilation=dilation,
groups=groups,
)
_release_backward_weight_slot(state, idx)
# Ensure previous grad transfer that used this slot is done
torch.cuda.current_stream().wait_event(ev_tx_w_bwd_done)
# Compute heavy grads on GPU into staging buffers
# Compute heavy grads on GPU into staging buffers (frozen bases skip this)
grad_weight = None
grad_bias = None
if (
need_w = (
getattr(weight_cpu, "requires_grad", False)
and weight_cpu.dtype.is_floating_point
):
w_grad_buffers[idx] = conv2d_weight(
x,
weight_cpu.shape,
grad_out,
stride=stride,
padding=padding,
dilation=dilation,
groups=groups,
)
need_b = bias_cpu is not None and getattr(bias_cpu, "requires_grad", False)
if need_w or need_b:
torch.cuda.current_stream().wait_event(state["grad_xfer_done"][idx])
w_grad_gpu = b_grad_gpu = None
if need_w:
w_grad_gpu = conv2d_weight(
x,
weight_cpu.shape,
grad_out,
stride=stride,
padding=padding,
dilation=dilation,
groups=groups,
)
state["w_grad_buffers"][idx] = w_grad_gpu
if need_b:
b_grad_gpu = grad_out.sum(dim=(0, 2, 3))
state["b_grad_buffers"][idx] = b_grad_gpu
grad_weight, grad_bias = _stage_grads_to_cpu(
state, idx, w_grad_gpu, b_grad_gpu
)
if bias_cpu is not None and getattr(bias_cpu, "requires_grad", False):
b_grad_buffers[idx] = grad_out.sum(dim=(0, 2, 3))
ev_cu_b_finish.record()
# Launch CPU copies on the dedicated grad stream (overlaps with next H2D)
with torch.cuda.stream(transfer_grad_stream):
transfer_grad_stream.wait_event(ev_cu_b_finish)
if (
getattr(weight_cpu, "requires_grad", False)
and weight_cpu.dtype.is_floating_point
):
grad_weight = w_grad_buffers[idx].to("cpu", non_blocking=True)
if bias_cpu is not None and getattr(bias_cpu, "requires_grad", False):
grad_bias = b_grad_buffers[idx].to("cpu", non_blocking=True)
state["transfer_weight_backward_finished_event"].record()
return (
grad_input.to(dtype=grad_out.dtype),