From 4257939e50739280203b0aec191efbcb7101a670 Mon Sep 17 00:00:00 2001
From: nimlgen <138685161+nimlgen@users.noreply.github.com>
Date: Tue, 14 Jul 2026 19:47:22 +0300
Subject: [PATCH] remove copyin/copyout from Buffer (#17020)

* remove copyin/copyout from Buffer

* x

* x

* x

* x
---
 extra/hcq2/hcq2.py                     | 19 -----------------
 test/backend/test_graph.py             | 11 ++++------
 test/backend/test_linearizer.py        |  2 +-
 test/backend/test_profiler.py          | 16 +++++++-------
 test/backend/test_renderer_failures.py |  6 +++---
 test/backend/test_subbuffer.py         | 23 ++++++++++----------
 test/backend/test_uops.py              | 14 ++++---------
 test/device/test_hcq.py                |  4 ++--
 test/device/test_ocl.py                |  4 ++--
 test/external/external_test_hcq.py     |  4 ++--
 test/external/fuzz_graph.py            | 10 ++++-----
 test/opt/test_tensor_cores.py          |  3 ++-
 test/unit/test_invalid_tensor.py       |  3 ++-
 tinygrad/device.py                     | 29 +++++++++++---------------
 tinygrad/engine/realize.py             |  5 ++---
 tinygrad/runtime/ops_python.py         |  9 +++++---
 tinygrad/runtime/support/hcq.py        |  5 +++--
 tinygrad/tensor.py                     |  2 +-
 tinygrad/uop/ops.py                    |  3 +--
 19 files changed, 70 insertions(+), 102 deletions(-)

diff --git a/extra/hcq2/hcq2.py b/extra/hcq2/hcq2.py
index 01d9901de5..1dc882f874 100644
--- a/extra/hcq2/hcq2.py
+++ b/extra/hcq2/hcq2.py
@@ -593,22 +593,3 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
     self.dev.iface.free(mb)
 
   def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
-
-  def _wrap(self, dev:str, sz:int, opaque:HCQ2Buffer) -> Buffer:
-    return Buffer(dev, sz, dtypes.uint8, opaque=opaque, options=BufferSpec(external_ptr=1))
-
-  def _copy(self, dst:Buffer, src:Buffer):
-    from tinygrad.engine.realize import run_linear
-    du, su = UOp.from_buffer(dst), UOp.from_buffer(src)
-    run_linear(UOp(Ops.LINEAR, src=(su.param_like(1).copy_to_device(dst.device).call(du, su),)), update_stats=True)
-
-  def _copyin(self, dest:HCQ2Buffer, src:memoryview):
-    s = Buffer(self.dev.device, len(src), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    s._buf.cpu_view()[:len(src)] = src
-    self._copy(self._wrap(self.dev.device, len(src), dest), s)
-
-  def _copyout(self, dest:memoryview, src:HCQ2Buffer):
-    d = Buffer(self.dev.device, len(dest), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    self._copy(d, self._wrap(self.dev.device, len(dest), src))
-    self.dev.synchronize()
-    dest[:] = d._buf.cpu_view()[:len(dest)]
diff --git a/test/backend/test_graph.py b/test/backend/test_graph.py
index 20fefe99c3..6d16dd5f2d 100644
--- a/test/backend/test_graph.py
+++ b/test/backend/test_graph.py
@@ -1,9 +1,9 @@
 import numpy as np
-import functools, unittest, ctypes
+import functools, unittest
 
 from tinygrad.device import Device, Buffer
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import Context, from_mv
+from tinygrad.helpers import Context
 from tinygrad.dtype import dtypes
 from tinygrad.engine.jit import MultiGraphRunner
 from tinygrad.engine.realize import run_linear, compile_linear
@@ -31,7 +31,7 @@ def make_buffer(device, size=BUF_SIZE, fill=False):
   buf = Buffer(device, size, dtypes.int).ensure_allocated()
   if fill:
     with Context(DEBUG=0):
-      buf.copyin(Tensor(np.random.randint(-10000, 10000, size=size, dtype=np.int32)).realize().uop.base.realized.as_memoryview())
+      buf.copy_from(Tensor(np.random.randint(-10000, 10000, size=size, dtype=np.int32)).realize().uop.base.realized)
   return buf
 
 def make_view(base, offset_elems, size_elems):
@@ -55,10 +55,7 @@ def run_schedule(calls:list[UOp]):
   run_linear(UOp(Ops.LINEAR, src=tuple(calls)))
 
 def zero_bufs(bufs):
-  for b in bufs:
-    mv = memoryview(bytearray(b.nbytes))
-    ctypes.memset(from_mv(mv), 0, len(mv))
-    b.copyin(mv)
+  for b in bufs: b.copy_from(Buffer("PYTHON", b.size, b.dtype, opaque=memoryview(bytearray(b.nbytes))))
 
 @unittest.skipUnless(Device[Device.DEFAULT].graph is not None, "graph support required")
 class TestGraph(unittest.TestCase):
diff --git a/test/backend/test_linearizer.py b/test/backend/test_linearizer.py
index cac2bd0ad0..cca3c7c7f8 100644
--- a/test/backend/test_linearizer.py
+++ b/test/backend/test_linearizer.py
@@ -424,7 +424,7 @@ def copyout_outputs(outbufs:list[Buffer]) -> list[np.ndarray]:
   return [np.frombuffer(x.as_memoryview(), _to_np_dtype(x.dtype)) for x in outbufs]
 
 def reset_bufs(bufs:list[Buffer]):
-  for buf in bufs: buf.copyin(np.zeros((buf.size*buf.dtype.itemsize,), dtype=np.uint8).data)
+  for buf in bufs: buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=memoryview(bytearray(buf.nbytes))))
 
 def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[],
                                apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[]):
diff --git a/test/backend/test_profiler.py b/test/backend/test_profiler.py
index 2091f85ac4..7c2796f7e1 100644
--- a/test/backend/test_profiler.py
+++ b/test/backend/test_profiler.py
@@ -70,9 +70,9 @@ class TestProfiler(unittest.TestCase):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
 
-    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
+    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith((TestProfiler.d0.device, "PYTHON"))]
     assert len(kernel_runs) == 1, "one kernel run is expected"
 
   def test_profile_multiops(self):
@@ -80,12 +80,12 @@ class TestProfiler(unittest.TestCase):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
       gs, ls = TestProfiler.prg.arg.launch_dims({})
       TestProfiler.runtime(buf1._buf, TestProfiler.a.uop.buffer._buf, global_size=gs, local_size=ls)
-      buf1.copyout(memoryview(bytearray(buf1.nbytes)))
+      buf1.as_memoryview()
 
-    evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
+    evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith((TestProfiler.d0.device, "PYTHON"))]
 
     assert len(evs) == 3, "3 kernel runs are expected"
     # NOTE: order of events does not matter, the tool is responsible for sorting them
@@ -103,12 +103,12 @@ class TestProfiler(unittest.TestCase):
     buf2 = Buffer(f"{Device.DEFAULT}:1", 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0, d1) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-      buf2.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+      buf2.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
 
     for dev in [TestProfiler.d0.device, d1.device]:
       evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and _dev_base(x.device) == dev]
-      assert len(evs) == 1, "one kernel runs are expected"
+      assert len(evs) == (0 if hasattr(TestProfiler.d0.allocator, '_as_buffer') else 1), "one kernel runs are expected"
 
   def test_profile_multidev_transfer(self):
     try: d1 = Device[f"{Device.DEFAULT}:1"]
diff --git a/test/backend/test_renderer_failures.py b/test/backend/test_renderer_failures.py
index 867980036b..b347dcc8d1 100644
--- a/test/backend/test_renderer_failures.py
+++ b/test/backend/test_renderer_failures.py
@@ -1,6 +1,6 @@
 import unittest
 import numpy as np
-from tinygrad.device import Device
+from tinygrad.device import Device, Buffer
 from tinygrad.dtype import dtypes, ConstType
 from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
@@ -10,13 +10,13 @@ from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.wgsl import WGSLRenderer
 from tinygrad.runtime.ops_python import PythonRenderer
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, python_alu
-from tinygrad.tensor import Tensor, _to_np_dtype
+from tinygrad.tensor import Tensor
 
 def _test_uop_result(inputs:list[Tensor], sink:UOp, local_size=None):
   for x in inputs: x.realize()
   sz = 1 if local_size is None else prod(local_size)
   outs = [UOp.new_buffer(Device.DEFAULT, sz, u.src[1].dtype) for u in sink.src if u.op is Ops.STORE]
-  for u in outs: u.buffer.allocate().copyin(np.zeros(sz, dtype=_to_np_dtype(u.dtype)).data)
+  for u in outs: u.buffer.allocate().copy_from(Buffer("PYTHON", sz, u.dtype, opaque=memoryview(bytearray(u.buffer.nbytes))))
   run_linear(UOp(Ops.LINEAR, src=(sink.call(*outs, *(x.uop.base for x in inputs)),)))
   return [u.buffer.numpy() for u in outs]
 
diff --git a/test/backend/test_subbuffer.py b/test/backend/test_subbuffer.py
index 8b1b3aef3c..69fdafdab1 100644
--- a/test/backend/test_subbuffer.py
+++ b/test/backend/test_subbuffer.py
@@ -7,8 +7,7 @@ from test.helpers import needs_second_gpu
 @unittest.skipIf(Device.DEFAULT in {"WEBGPU", "CL"}, "subbuffer not supported")
 class TestSubBuffer(unittest.TestCase):
   def setUp(self):
-    self.buf = Buffer(Device.DEFAULT, 10, dtypes.uint8).ensure_allocated()
-    self.buf.copyin(memoryview(bytearray(range(10))))
+    self.buf = Buffer(Device.DEFAULT, 10, dtypes.uint8, initial_value=bytes(range(10)))
     self.buf_unalloc = Buffer(Device.DEFAULT, 10, dtypes.uint8)
 
   def test_subbuffer(self):
@@ -59,7 +58,7 @@ class TestSubBuffer(unittest.TestCase):
       _ = Buffer(Device.DEFAULT, 10, dtypes.uint8).ensure_allocated()
 
       self.buf.ensure_allocated()
-      self.buf.copyin(memoryview(bytearray(range(10, 20))))
+      self.buf.copy_from(Buffer("PYTHON", 10, dtypes.uint8, opaque=memoryview(bytearray(range(10, 20)))))
 
       vbuf.ensure_allocated()
 
@@ -109,15 +108,15 @@ class TestSubBuffer(unittest.TestCase):
   def test_subbuffer_copy_in_out(self):
     sub_buf = self.buf.view(3, dtypes.uint8, offset=3).ensure_allocated() # [3:6]
     data_out_sub = bytearray([0]*3)
-    sub_buf.copyout(memoryview(data_out_sub))
+    data_out_sub[:] = sub_buf.as_memoryview()
     assert data_out_sub == bytearray(range(3, 6))
-    sub_buf.copyin(memoryview(bytearray(range(3))))
+    sub_buf.copy_from(Buffer("PYTHON", 3, dtypes.uint8, opaque=memoryview(bytearray(range(3)))))
     assert sub_buf.as_memoryview().tolist() == list(range(3))
     assert self.buf.as_memoryview().tolist()[3:6] == list(range(3))
-    sub_buf.copyout(memoryview(data_out_sub))
+    data_out_sub[:] = sub_buf.as_memoryview()
     assert data_out_sub == bytearray(range(3))
     data_out_base = bytearray([0]*10)
-    self.buf.copyout(memoryview(data_out_base))
+    data_out_base[:] = self.buf.as_memoryview()
     assert data_out_base[0:3] == bytearray(range(0, 3))
     assert data_out_base[3:6] == data_out_sub
     assert data_out_base[6:10] == bytearray(range(6, 10))
@@ -129,27 +128,27 @@ class TestSubBuffer(unittest.TestCase):
     self.assertTrue(view2.is_allocated())
 
     data_in = bytearray([7, 8, 9])
-    view2.copyin(memoryview(data_in))
+    view2.copy_from(Buffer("PYTHON", 3, view2.dtype, opaque=memoryview(data_in)))
     data_out_v2 = bytearray([0]*3)
-    view2.copyout(memoryview(data_out_v2))
+    data_out_v2[:] = view2.as_memoryview()
     assert data_in == data_out_v2
 
     expected_base_data = memoryview(bytearray(range(10)))
     expected_base_data[4:7] = data_in
 
     data_out_base = bytearray([0]*10)
-    self.buf.copyout(memoryview(data_out_base))
+    data_out_base[:] = self.buf.as_memoryview()
     assert expected_base_data == data_out_base
 
   def test_subbuffer_alloc(self):
     sub_buf = self.buf.view(4, dtypes.int8, offset=3)
     sub_buf.allocate()
-    sub_buf.copyin(memoryview(bytearray(range(10, 14))))
+    sub_buf.copy_from(Buffer("PYTHON", 4, dtypes.int8, opaque=memoryview(bytearray(range(10, 14)))))
     assert self.buf.as_memoryview().tolist()[3:7] == sub_buf.as_memoryview().tolist()
 
     sub_buf = self.buf_unalloc.view(4, dtypes.int8, offset=3)
     sub_buf.allocate()
-    sub_buf.copyin(memoryview(bytearray(range(10, 14))))
+    sub_buf.copy_from(Buffer("PYTHON", 4, dtypes.int8, opaque=memoryview(bytearray(range(10, 14)))))
     assert self.buf_unalloc.as_memoryview().tolist()[3:7] == sub_buf.as_memoryview().tolist()
 
   def test_subbuffer_dealloc(self):
diff --git a/test/backend/test_uops.py b/test/backend/test_uops.py
index 31a7647a51..af1164fb1b 100644
--- a/test/backend/test_uops.py
+++ b/test/backend/test_uops.py
@@ -33,11 +33,9 @@ def _test_single_value(vals, op, dts):
   alu = uop(uops, op, output_dtype, loads)
   out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), alu))
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
-  buf2 = [Buffer(Device.DEFAULT, 1, dtype).allocate().copyin(np.array([a], dtype=_to_np_dtype(dtype)).data) for a,dtype in zip(vals, dts)]
+  buf2 = [Buffer(Device.DEFAULT, 1, dtype, initial_value=np.array([a], dtype=_to_np_dtype(dtype)).tobytes()) for a,dtype in zip(vals, dts)]
   run_uops([out], [buf]+buf2)
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 def _test_single_value_const(vals, op, dts):
   uops = []
@@ -48,9 +46,7 @@ def _test_single_value_const(vals, op, dts):
   out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   run_uops([out], [buf])
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 def _test_uops_result(output_dtype, uops, res):
   # uops = []
@@ -59,9 +55,7 @@ def _test_uops_result(output_dtype, uops, res):
   out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), res))
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   run_uops([out], [buf])
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 class TestUOps(unittest.TestCase):
   def _equal(self, v1, v2):
diff --git a/test/device/test_hcq.py b/test/device/test_hcq.py
index e0d4fe32e4..369adddf8d 100644
--- a/test/device/test_hcq.py
+++ b/test/device/test_hcq.py
@@ -31,8 +31,8 @@ class TestHCQ(unittest.TestCase):
 
   def setUp(self):
     TestHCQ.d0.synchronize()
-    TestHCQ.a.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-    TestHCQ.b.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 0))))
+    TestHCQ.a.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+    TestHCQ.b.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 0)))))
     TestHCQ.d0.synchronize() # wait for copyins to complete
 
   # Test signals
diff --git a/test/device/test_ocl.py b/test/device/test_ocl.py
index 3923a95093..90bfbe55e4 100644
--- a/test/device/test_ocl.py
+++ b/test/device/test_ocl.py
@@ -34,7 +34,7 @@ class TestCLError(unittest.TestCase):
     data = list(range(65))
     unaligned = memoryview(bytearray(data))[1:]
     buffer = Buffer("CL", 64, dtypes.uint8).allocate()
-    buffer.copyin(unaligned)
+    buffer.copy_from(Buffer("PYTHON", 64, dtypes.uint8, opaque=unaligned))
     result = memoryview(bytearray(len(data) - 1))
-    buffer.copyout(result)
+    result[:] = buffer.as_memoryview()
     assert unaligned == result, "Unaligned data copied in must be equal to data copied out."
diff --git a/test/external/external_test_hcq.py b/test/external/external_test_hcq.py
index cb21563ae1..ec65fffd7e 100644
--- a/test/external/external_test_hcq.py
+++ b/test/external/external_test_hcq.py
@@ -47,8 +47,8 @@ class TestHCQ(unittest.TestCase):
 
   def setUp(self):
     TestHCQ.d0.synchronize()
-    TestHCQ.a.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-    TestHCQ.b.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 0))))
+    TestHCQ.a.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+    TestHCQ.b.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 0)))))
     TestHCQ.d0.synchronize() # wait for copyins to complete
 
   def test_run_1000_times_one_submit(self):
diff --git a/test/external/fuzz_graph.py b/test/external/fuzz_graph.py
index edc2c6faee..e2a6877298 100644
--- a/test/external/fuzz_graph.py
+++ b/test/external/fuzz_graph.py
@@ -1,7 +1,7 @@
-import random, ctypes
+import random
 import numpy as np
 from tinygrad.device import Buffer, Device
-from tinygrad.helpers import Context, getenv, from_mv
+from tinygrad.helpers import Context, getenv
 from tinygrad.dtype import dtypes
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import BufferXfer, get_runner, ExecItem
@@ -29,7 +29,7 @@ def alloc_rawbuffer(device, fill=False):
   if fill:
     with Context(DEBUG=0):
       data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
+      rawbuf.copy_from(Tensor(data).realize().uop.base.realized)
   return rawbuf
 
 def gen_kernel_ji(device, deps):
@@ -84,9 +84,7 @@ def run_jit(jis, all_buffers, input_buffers, var_vals):
   with Context(DEBUG=0):
     for rawbuf in all_buffers:
       if rawbuf in input_buffers: continue
-      mv = memoryview(bytearray(rawbuf.nbytes))
-      ctypes.memset(from_mv(mv), 0, len(mv))
-      rawbuf.copyin(mv)
+      rawbuf.copy_from(Buffer("PYTHON", rawbuf.size, rawbuf.dtype, opaque=memoryview(bytearray(rawbuf.nbytes))))
 
   for ei in jis: ei.run(var_vals, jit=True)
 
diff --git a/test/opt/test_tensor_cores.py b/test/opt/test_tensor_cores.py
index 9c269f1f6d..d9595fe0b8 100644
--- a/test/opt/test_tensor_cores.py
+++ b/test/opt/test_tensor_cores.py
@@ -149,7 +149,8 @@ class TestTensorCores(unittest.TestCase):
 
         # TODO: support this even if numpy doesn't
         if _to_np_dtype(real_bufs[0].dtype) is None: continue
-        real_bufs[0].copyin(np.zeros((real_bufs[0].size, ), dtype=_to_np_dtype(real_bufs[0].dtype)).data) # Zero to check that all values are filled
+        # Zero to check that all values are filled
+        real_bufs[0].copy_from(Buffer("PYTHON", real_bufs[0].size, real_bufs[0].dtype, opaque=memoryview(bytearray(real_bufs[0].nbytes))))
         run_program(ast, real_bufs)
         result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
 
diff --git a/test/unit/test_invalid_tensor.py b/test/unit/test_invalid_tensor.py
index 2e051cbac9..14b6a8570b 100644
--- a/test/unit/test_invalid_tensor.py
+++ b/test/unit/test_invalid_tensor.py
@@ -1,5 +1,6 @@
 import unittest
 from tinygrad import Tensor
+from tinygrad.device import Buffer
 from tinygrad.dtype import Invalid, dtypes
 from tinygrad.engine.realize import run_linear
 
@@ -9,7 +10,7 @@ class TestInvalidTensor(unittest.TestCase):
     buf = out.uop.buffer
     buf.allocate()
     sentinel = memoryview(bytearray(b'\x42' * buf.nbytes))
-    buf.copyin(sentinel)
+    buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=sentinel))
     before = buf.as_memoryview().cast(out.dtype.fmt).tolist()
     run_linear(linear, var_vals)
     ret = buf.as_memoryview().cast(out.dtype.fmt).tolist()
diff --git a/tinygrad/device.py b/tinygrad/device.py
index b53e65faa1..f3cdd0855d 100644
--- a/tinygrad/device.py
+++ b/tinygrad/device.py
@@ -3,7 +3,7 @@ from dataclasses import dataclass, replace
 from collections import defaultdict
 from typing import Any, Generic, TypeVar, Iterator, Generator, TYPE_CHECKING
 import importlib, inspect, functools, pathlib, os, contextlib, re, atexit, pickle, decimal
-from tinygrad.helpers import LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored
+from tinygrad.helpers import LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, PROFILE, temp, colored
 from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, suppress_finalizing
 from tinygrad.helpers import select_by_name, select_first_inited, DEV, TracingKey, size_to_str, pluralize
 from tinygrad.dtype import DType, _to_np_dtype
@@ -112,7 +112,7 @@ class Buffer:
       if opaque is not None: self.allocate(opaque)
       if initial_value is not None:
         self.allocate()
-        self.copyin(memoryview(initial_value))
+        self.copy_from(Buffer("PYTHON", self.size, self.dtype, opaque=memoryview(bytearray(initial_value))))
     else:
       assert base._base is None, "base can't have a base"
       assert device == base.device, "base must have the same device"
@@ -176,9 +176,7 @@ class Buffer:
     if self._base is not None:
       return self.__class__, (self.device, self.size, self.dtype, None, None, None, 0, self.base, self.offset, self.is_allocated())
     if self.device == "NPY": return self.__class__, (self.device, self.size, self.dtype, self._buf, self.options, None, self.uop_refcount)
-    if self.is_allocated():
-      buf = bytearray(self.nbytes)
-      self.copyout(memoryview(buf))
+    if self.is_allocated(): buf = bytearray(self.as_memoryview())
     return self.__class__, (self.device, self.size, self.dtype, None, self.options, buf, self.uop_refcount)
   @property
   def trace_num(self) -> int:
@@ -195,23 +193,20 @@ class Buffer:
     # zero copy with as_memoryview (disabled by default due to use after free)
     if (force_zero_copy or allow_zero_copy) and hasattr(self.allocator, '_as_buffer'): return self.allocator._as_buffer(self._buf)
     assert not force_zero_copy, "force zero copy was passed, but copy is required"
-    return self.copyout(memoryview(bytearray(self.nbytes)))
+    Buffer("PYTHON", self.size, self.dtype, opaque=(mv:=memoryview(bytearray(self.nbytes)))).copy_from(self)
+    return mv
   def numpy(self) -> 'np.ndarray': # type: ignore [name-defined] # noqa: F821
     import numpy as np
     assert _to_np_dtype(self.dtype) is not None, f"no np dtype for {self.dtype}"
     return np.frombuffer(self.as_memoryview(), dtype=_to_np_dtype(self.dtype))
-  def copyin(self, mv:memoryview):
-    mv = flat_mv(mv)
-    assert len(mv) == self.nbytes, f"size mismatch, {len(mv)=} != {self.dtype=} {self.size=}"
-    assert self.is_initialized(), "can't copyin to unallocated buffer"
-    self.allocator._copyin(self._buf, mv)
+  def copy_from(self, src:Buffer) -> Buffer:
+    assert self.nbytes == src.nbytes, f"copy size mismatch, {self.nbytes} != {src.nbytes}"
+    assert self.is_initialized() and src.is_initialized(), "copy requires allocated buffers"
+    from tinygrad.engine.realize import run_linear
+    from tinygrad.uop.ops import UOp, Ops
+    du, su = UOp.from_buffer(self), UOp.from_buffer(src)
+    run_linear(UOp(Ops.LINEAR, src=(su.param_like(1).copy_to_device(self.device).call(du, su),)), update_stats=False)
     return self
-  def copyout(self, mv:memoryview) -> memoryview:
-    mv = flat_mv(mv)
-    assert len(mv) == self.nbytes, f"size mismatch, {len(mv)=} != {self.dtype=} {self.size=}"
-    assert self.is_initialized(), "can't copyout unallocated buffer"
-    self.allocator._copyout(mv, self._buf)
-    return mv
   def view(self, size:int, dtype:DType, offset:int) -> Buffer:
     assert offset < self.nbytes, "offset must be less than nbytes"
     return Buffer(self.device, size, dtype, base=self.base, offset=self.offset+offset)
diff --git a/tinygrad/engine/realize.py b/tinygrad/engine/realize.py
index 0b68e4100d..3a64d7b0bf 100644
--- a/tinygrad/engine/realize.py
+++ b/tinygrad/engine/realize.py
@@ -166,9 +166,8 @@ def exec_copy(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
       elif src.device.startswith("DISK") and getattr(src.allocator.dev, 'fd', None) is not None \
            and hasattr(dest.allocator, 'copy_from_disk') and src.nbytes >= 4096 and dest.allocator.supports_copy_from_disk:
         dest.allocator.copy_from_disk(dest._buf, src._buf, src.nbytes)
-      elif src.device.startswith(("DISK", "TINYFS")) and hasattr(dest.allocator, '_as_buffer'):
-        src.allocator._copyout(dest.allocator._as_buffer(dest._buf), src._buf)
-      else: dest.copyin(src.as_memoryview(allow_zero_copy=True))
+      elif hasattr(dest.allocator, '_as_buffer'): src.allocator._copyout(dest.allocator._as_buffer(dest._buf), src._buf)
+      else: dest.allocator._copyin(dest._buf, src.as_memoryview(allow_zero_copy=True))
   return None
 
 def exec_kernel(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
diff --git a/tinygrad/runtime/ops_python.py b/tinygrad/runtime/ops_python.py
index 766ae734d1..44def8a56d 100644
--- a/tinygrad/runtime/ops_python.py
+++ b/tinygrad/runtime/ops_python.py
@@ -6,7 +6,7 @@ from typing import Any, TYPE_CHECKING
 import pickle, base64, itertools, time, sys, functools
 from dataclasses import replace
 from tinygrad.dtype import DType, dtypes, AddrSpace, truncate, storage_fmt_for_dtype, to_storage_scalar, from_storage_scalar
-from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape
+from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape, cpu_profile
 from tinygrad.device import Buffer, Compiled, Compiler, Allocator
 from tinygrad.codegen.opt import tc
 from tinygrad.uop.ops import exec_alu, python_alu, Ops, UOp, GroupOp, bitcast
@@ -223,8 +223,11 @@ class PythonRenderer(Renderer):
 
 class PythonAllocator(Allocator['PythonDevice']):
   def _alloc(self, size, options): return memoryview(bytearray(size))
-  def _copyin(self, dest, src:memoryview): dest[:] = src
-  def _copyout(self, dest:memoryview, src): dest[:] = src
+  def _as_buffer(self, src) -> memoryview: return src
+  def _copyin(self, dest, src:memoryview):
+    with cpu_profile("TINY -> PYTHON", f"{self.dev.device}:COPY"): dest[:] = src
+  def _copyout(self, dest:memoryview, src):
+    with cpu_profile("PYTHON -> TINY", f"{self.dev.device}:COPY"): dest[:] = src
   def map(self, buf:Buffer): return buf.as_memoryview(force_zero_copy=True)
   def _offset(self, buf:memoryview, size:int, offset:int): return buf[offset:offset+size]
 
diff --git a/tinygrad/runtime/support/hcq.py b/tinygrad/runtime/support/hcq.py
index 62cce31367..f302eaf4d3 100644
--- a/tinygrad/runtime/support/hcq.py
+++ b/tinygrad/runtime/support/hcq.py
@@ -551,8 +551,8 @@ class HCQAllocatorBase(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
     self.b = copy_bufs or [self._alloc(batch_size, BufferSpec(host=True)) for _ in range(batch_cnt)]
     self.b_timeline, self.b_next, self.max_copyout_size = [0] * len(self.b), 0, max_copyout_size
 
-  def _map(self, buf:HCQBuffer):
-    if self.dev in buf.mapped_devs: return
+  def _map(self, buf:HCQBuffer) -> HCQBuffer:
+    if self.dev in buf.mapped_devs: return buf
     if buf.owner is None: raise RuntimeError(f"map failed: buffer {buf.va_addr} has no owner, it's a virtual buffer")
     if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
 
@@ -560,6 +560,7 @@ class HCQAllocatorBase(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
     # Devices can save mappings and internal metadata as a new buffer.
     if (mb:=self._do_map(buf)) is not None: buf.mappings[self.dev] = mb
     buf.mapped_devs.append(self.dev)
+    return buf
 
   @suppress_finalizing
   def _free(self, buf:HCQBuffer, options:BufferSpec|None=None):
diff --git a/tinygrad/tensor.py b/tinygrad/tensor.py
index 0f10a0cda4..5da95e585a 100644
--- a/tinygrad/tensor.py
+++ b/tinygrad/tensor.py
@@ -218,7 +218,7 @@ class Tensor(RandMixin):
 
     # TODO: this is a hack for writing to DISK. remove with working assign
     if is_disk:
-      self._buffer().copyin(x._data())
+      (b:=self._buffer()).copy_from(Buffer("PYTHON", b.size, b.dtype, opaque=x._data()))
       return self
     # STORE+AFTER: STORE is the write effect (void), AFTER wraps the view for correct shape/ranging
     assign = self.uop.after(self.uop.store(x.uop))
diff --git a/tinygrad/uop/ops.py b/tinygrad/uop/ops.py
index 56458c0e25..56e3e511fc 100644
--- a/tinygrad/uop/ops.py
+++ b/tinygrad/uop/ops.py
@@ -770,8 +770,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
       assert bdtype.fmt is not None, f"{bdtype=} has None fmt"
       ret = UOp.empty(shape:=get_shape(x), dtype=bdtype, device="PYTHON")
       data = struct.pack(f"{prod(shape)}{bdtype.fmt}", *[truncate[bdtype](bdtype.const(xi)) for xi in fully_flatten(x)])
-    # fake realize. if target device is PYTHON it needs bytearray to be writable
-    ret.buffer.allocate(memoryview(data if device != "PYTHON" else bytearray(data)))
+    ret.buffer.allocate(memoryview(bytearray(data))) # fake realize. buffer storage must be writable, and bytes isn't
     if ret.dtype != dtype: ret = ret.cast(dtype)
     return ret if ret.device == device else ret.copy_to_device(device)
   def clone(self, device=None) -> UOp: