fix beam

* python bf16

* _to_torch_storage_type

---------

Co-authored-by: b1tg <[email protected]>

.github/actions/setup-tinygrad/action.yml

@@ -225,13 +225,17 @@ runs:
     - name: Install gpuocelot dependencies (MacOS)
       if: inputs.ocelot == 'true' && runner.os == 'macOS'
       shell: bash
-      run: brew install --quiet cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
+      run: |
+        pkgs=(cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses)
+        for f in "${pkgs[@]}"; do
+          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
+        done
     - name: Cache gpuocelot
       if: inputs.ocelot == 'true'
       id: cache-build
       uses: actions/cache@v4
       env:
-        cache-name: cache-gpuocelot-build
+        cache-name: cache-gpuocelot-build-1
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
         key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}

.github/workflows/test.yml

@@ -343,6 +343,8 @@ jobs:
       run: |
         python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
+    - name: Run TYPED=1
+      run: TYPED=1 python -c "import tinygrad"
 
   unittest:
     name: Unit Tests

docs/tinybox.md

@@ -6,7 +6,7 @@ If you don't have a tinybox and you want one, see [tinygrad.org](https://tinygra
 
 ## Welcome
 
-Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, and the green box includes six 4090 GPUs. Whether you bought a red one or a green one, we want you to love it.
+Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, the green box includes six 4090 GPUs, and the green v2 box includes four 5090 GPUs. Whether you bought a red one or a green one, we want you to love it.
 
 We don't have a stupid cloud service, you don't have to create a tiny account to set it up, and we aren't tracking how you use the box. We're just happy you bought one. This petaflop is your petaflop.
 

extra/amdpci/proclogs.py

@@ -37,7 +37,7 @@ def main():
   dev = PCIIface(None, 0)
   for x, y in dev.dev_impl.__dict__.items():
     if isinstance(y, AMRegister):
-      for inst, addr in y.addr.keys(): reg_names[addr] = f"{x}, xcc={inst}"
+      for inst, addr in y.addr.items(): reg_names[addr] = f"{x}, xcc={inst}"
 
   with open(sys.argv[1], 'r') as f:
     log_content = log_content_them = f.read()

extra/gemm/amd_uop_matmul.py

@@ -5,7 +5,7 @@ from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv, colored, prod, unwrap
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors
+from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
 from tinygrad.codegen.opt.swizzler import merge_views, view_left
 
 def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])
@@ -44,13 +44,28 @@ pm = PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
 ])
 
+def rangeify_kernel3():
+  a = Tensor.empty(N,N)
+  b = Tensor.empty(N,N)
+  c = a@b
+  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
+  with Context(RANGEIFY=1):
+    sink = c.schedule()[-1].ast
+  #print(sink)
+
+  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
+  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
+  opts += [Opt(OptOps.UNROLL, 0, 8)]
+
+  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
+
 def top_spec_kernel3():
   a = Tensor.empty(N,N)
   b = Tensor.empty(N,N)
   c = a@b
   sink = c.schedule()[-1].ast
   L = 16
-  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(dtypes.int, N//BM, 0), 2:UOp.range(dtypes.int, N//BN, 1)})
+  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(N//BM, 0), 2:UOp.range(N//BN, 1)})
   sink = graph_rewrite(sink, view_left+pm)
   axis_types = (AxisType.GLOBAL, AxisType.LOCAL, AxisType.GLOBAL, AxisType.LOCAL, AxisType.REDUCE)
   return sink.replace(arg=KernelInfo(name="top_"+to_colored(sink.full_shape, axis_types), axis_types=axis_types))
@@ -171,7 +186,7 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
 
   c_regs = UOp(Ops.DEFINE_REG, dtypes.float.ptr(TM * nbIterWaveM * TN * nbIterWaveN), arg=2)
 
-  i = UOp.range(dtypes.int, c_regs.dtype.size, 16)
+  i = UOp.range(c_regs.dtype.size, 16)
   init_store = c_regs[i].store(UOp.const(dtypes.float, 0.0), i)
 
   if kernel4:
@@ -182,53 +197,53 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     kId = 0
 
     # load from globals into locals
-    i = UOp.range(dtypes.int, nbReadsB, 0)
+    i = UOp.range(nbReadsB, 0)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 1)
+    i = UOp.range(nbReadsA, 1)
     index_x = rAIdx + kId
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
 
     # iterate over the middle chunk
-    kId_range = UOp.range(dtypes.int, N//BK-1, 2)
+    kId_range = UOp.range(N//BK-1, 2)
     kId = kId_range*BK
 
     barrier = UOp.barrier(As_store, Bs_store)
 
     # load from globals into registers (next round)
-    i = UOp.range(dtypes.int, nbReadsB, 3)
+    i = UOp.range(nbReadsB, 3)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId + BK
     regB_store = regB[i].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 4)
+    i = UOp.range(nbReadsA, 4)
     index_x = rAIdx + kId + BK
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     regA_store = regA[i].store(a[N * index_y + index_x].load(), i)
 
     def inner_loop(first_range, inp_dep=()):
       # inner unroll
-      k = UOp.range(dtypes.int, BK, first_range+0)
+      k = UOp.range(BK, first_range+0)
 
       # load from locals into registers
-      iterWave = UOp.range(dtypes.int, nbIterWaveN, first_range+1)
-      i = UOp.range(dtypes.int, TN, first_range+2)
+      iterWave = UOp.range(nbIterWaveN, first_range+1)
+      i = UOp.range(TN, first_range+2)
       index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
       B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(*inp_dep), iterWave, i)
 
-      iterWave = UOp.range(dtypes.int, nbIterWaveM, first_range+3)
-      i = UOp.range(dtypes.int, TM, first_range+4)
+      iterWave = UOp.range(nbIterWaveM, first_range+3)
+      i = UOp.range(TM, first_range+4)
       index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
       A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(*inp_dep), iterWave, i)
 
       # do the GEMM math
-      iterWaveM = UOp.range(dtypes.int, nbIterWaveM, first_range+5)
-      yt = UOp.range(dtypes.int, TM, first_range+6)
-      iterWaveN = UOp.range(dtypes.int, nbIterWaveN, first_range+7)
-      xt = UOp.range(dtypes.int, TN, first_range+8)
+      iterWaveM = UOp.range(nbIterWaveM, first_range+5)
+      yt = UOp.range(TM, first_range+6)
+      iterWaveN = UOp.range(nbIterWaveN, first_range+7)
+      xt = UOp.range(TN, first_range+8)
       x = iterWaveN * TN + xt
       y = iterWaveM * TM + yt
       c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -241,12 +256,12 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     sink = inner_loop(5, (barrier, regB_store, regA_store)).barrier()
 
     # load from registers into locals
-    i = UOp.range(dtypes.int, nbReadsB, 14)
+    i = UOp.range(nbReadsB, 14)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId + BK
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(regB[i].load(sink), i, kId_range)
 
-    i = UOp.range(dtypes.int, nbReadsA, 15)
+    i = UOp.range(nbReadsA, 15)
     index_x = rAIdx + kId + BK
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(regA[i].load(sink), i, kId_range)
@@ -254,40 +269,40 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
     # final iteration without the copy
     sink = inner_loop(16, (UOp.barrier(Bs_store, As_store),))
   else:
-    kId_range = UOp.range(dtypes.int, N//BK, 0)
+    kId_range = UOp.range(N//BK, 0)
     kId = kId_range*BK
 
     # load from globals into locals
-    i = UOp.range(dtypes.int, nbReadsB, 1)
+    i = UOp.range(nbReadsB, 1)
     index_x = BN * blockIdx_x + rBIdx
     index_y = rBIdy + i * strideReadB + kId
     Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
 
-    i = UOp.range(dtypes.int, nbReadsA, 2)
+    i = UOp.range(nbReadsA, 2)
     index_x = rAIdx + kId
     index_y = BM * blockIdx_y + rAIdy + i * strideReadA
     As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
 
     barrier = UOp.barrier(As_store, Bs_store)
 
-    k = UOp.range(dtypes.int, BK, 3)
+    k = UOp.range(BK, 3)
 
     # load from locals into registers
-    iterWave = UOp.range(dtypes.int, nbIterWaveN, 4)
-    i = UOp.range(dtypes.int, TN, 5)
+    iterWave = UOp.range(nbIterWaveN, 4)
+    i = UOp.range(TN, 5)
     index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
     B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(barrier), iterWave, i)
 
-    iterWave = UOp.range(dtypes.int, nbIterWaveM, 6)
-    i = UOp.range(dtypes.int, TM, 7)
+    iterWave = UOp.range(nbIterWaveM, 6)
+    i = UOp.range(TM, 7)
     index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
     A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(barrier), iterWave, i)
 
     # do the GEMM math
-    iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 8)
-    yt = UOp.range(dtypes.int, TM, 9)
-    iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 10)
-    xt = UOp.range(dtypes.int, TN, 12)
+    iterWaveM = UOp.range(nbIterWaveM, 8)
+    yt = UOp.range(TM, 9)
+    iterWaveN = UOp.range(nbIterWaveN, 10)
+    xt = UOp.range(TN, 12)
     x = iterWaveN * TN + xt
     y = iterWaveM * TM + yt
     c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -295,10 +310,10 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
                             iterWaveM, iterWaveN, yt, xt, k, kId_range)
 
   # store c_regs into c
-  iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 1000)
-  yt = UOp.range(dtypes.int, TM, 1001)
-  iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 1002)
-  xt = UOp.range(dtypes.int, TN, 1003)
+  iterWaveM = UOp.range(nbIterWaveM, 1000)
+  yt = UOp.range(TM, 1001)
+  iterWaveN = UOp.range(nbIterWaveN, 1002)
+  xt = UOp.range(TN, 1003)
   xOut = blockIdx_x * BN + waveIdx * WN + iterWaveN * SUBWN + TN * idxInWave
   yOut = blockIdx_y * BM + waveIdy * WM + iterWaveM * SUBWM + TM * idyInWave
   indexC = N * (yOut + yt) + xOut + xt
@@ -309,10 +324,15 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
 
 if __name__ == "__main__":
   HL = getenv("HL")
-  if HL == 2: hprg = top_spec_kernel3()
+  if HL == 3: hprg = rangeify_kernel3()
+  elif HL == 2: hprg = top_spec_kernel3()
   elif HL == 1: hprg = hl_spec_kernel3()
   else: hprg = hand_spec_kernel3()
-  prg = get_program(hprg, Device.default.renderer)
+  if HL == 3:
+    with Context(RANGEIFY=1, BLOCK_REORDER=0):
+      prg = get_program(hprg, Device.default.renderer)
+  else:
+    prg = get_program(hprg, Device.default.renderer)
   print(prg.src)
   if getenv("SRC"): exit(0)
   hrunner = CompiledRunner(prg)

extra/gemm/max_matmul.py

@@ -56,7 +56,7 @@ def randoms():
 def ast_to_cuda_prog(compiler, ast, opts):
   k = Kernel(ast)
   k.apply_opts(opts)
-  p = get_program(k.get_optimized_ast(), k.opts)
+  p = get_program(k.ast, k.opts, k.applied_opts)
   return CUDAProgram(device, p.function_name, compiler.compile(p.src))
 
 if __name__ == "__main__":

extra/gemm/tinygrad_nv_matmul.py

@@ -29,7 +29,7 @@ if __name__ == "__main__":
       Opt(op=OptOps.LOCAL, axis=0, amt=2),
     ]
   k.apply_opts(opts)
-  prg = get_program(k.get_optimized_ast(), k.opts)
+  prg = get_program(k.ast, k.opts, k.applied_opts)
   new_src = prg.src
   # can mod source here
   prg = replace(prg, src=new_src)

extra/optimization/test_beam_search.py

@@ -16,9 +16,9 @@ class TestBeamSearch(unittest.TestCase):
     BEAM.value = self.old_beam
 
   def test_variable_ast_beam(self):
-    with Context(IGNORE_OOB=1):
-      a = rand(3, 3).reshape((Variable("a", 1, 10).bind(3), 3))
-      a = (a+1).realize()
+    vi = Variable("a", 1, 10).bind(3)
+    a = rand(10, 3)[:vi]
+    a = (a+1).realize()
 
   def test_big_prime_number(self):
     a = rand(367, 367)
@@ -42,18 +42,16 @@ class TestBeamSearch(unittest.TestCase):
 
   def test_variable_big_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
-    a = rand(367, 367)
-    b = rand(367, 367)
-    with Context(IGNORE_OOB=1):
-      c = (a.reshape(367, v) @ b.reshape(v, 367)).realize()
-      np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), atol=1e-4, rtol=1e-4)
+    a = rand(367, 400)
+    b = rand(400, 367)
+    c = (a[:, :v] @ b[:v, :]).realize()
+    np.testing.assert_allclose(c.numpy(), a[:, :367].numpy() @ b[:367, :].numpy(), atol=1e-4, rtol=1e-4)
 
   def test_variable_shrink_prime_number(self):
     v = Variable("v", 1, 400).bind(367)
     a = rand(400, 367)
-    with Context(IGNORE_OOB=1):
-      b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
-      np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
+    b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
+    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
 
   def test_no_mutate_rawbuffers(self):
     a = rand(3, 3).realize()

extra/replay_pkl.py

@@ -58,7 +58,7 @@ if __name__ == "__main__":
             GlobalCounters.kernel_count -= 1
 
         if not getenv("NOOPT"): k.apply_opts(hand_coded_optimizations(k))
-        p2 = get_program(k.get_optimized_ast(), k.opts)
+        p2 = get_program(k.ast, k.opts, k.applied_opts)
         new_ei = replace(ei, prg=CompiledRunner(p2))
         new_ei.run()
         new_jit.append(new_ei)

setup.py

@@ -64,6 +64,7 @@ setup(name='tinygrad',
             "pre-commit",
             "ruff",
             "numpy",
+            "typeguard",
         ],
         #'mlperf': ["mlperf-logging @ git+https://github.com/mlperf/[email protected]"],
         'testing_minimal': testing_minimal,

test/external/external_benchmark_sdxl_softmax.py

@@ -24,5 +24,5 @@ if __name__ == "__main__":
   #k.apply_opt(Opt(OptOps.GROUP, 1, 32))
   #k.apply_opt(Opt(OptOps.GROUP, 0, 32))
   from tinygrad.engine.realize import CompiledRunner, ExecItem
-  run = CompiledRunner(prg:=get_program(k.get_optimized_ast(), k.opts))
+  run = CompiledRunner(prg:=get_program(k.ast, k.opts, k.applied_opts))
   ExecItem(run, si.bufs).run()

test/external/external_debug_metal_sd_conv.py

@@ -35,7 +35,7 @@ k = Kernel(ast)
 k.apply_opts(opts)
 bufs = bufs_from_lin(k)
 
-prg = CompiledRunner(get_program(k.get_optimized_ast(), k.opts))
+prg = CompiledRunner(get_program(k.ast, k.opts, k.applied_opts))
 
 for i in range(10):
   speed = prg(bufs, var_vals={}, wait=True)

test/external/external_test_onnx_backend.py

@@ -134,7 +134,6 @@ backend_test.exclude('test_simple_rnn_*')
 
 # no control flow
 # control flow uses AttributeProto.GRAPH
-backend_test.exclude('test_if_*')
 backend_test.exclude('test_loop*')
 backend_test.exclude('test_range_float_type_positive_delta_expanded_cpu') # requires loop
 backend_test.exclude('test_affine_grid_2d_align_corners_expanded_cpu')
@@ -183,6 +182,8 @@ backend_test.exclude('test_resize_downsample_scales_cubic_antialias_cpu') # anti
 backend_test.exclude('test_resize_downsample_sizes_cubic_antialias_cpu') # antialias not implemented
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_value_only_mapping_cpu') # bad data type string
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_mapping_cpu') # bad data type string
+backend_test.exclude('test_if_opt_cpu') # ValueError: 13 is not a valid AttributeType
+backend_test.exclude('test_if_seq_cpu') # NotImplementedError: op='SequenceConstruct' is not supported
 
 backend_test.exclude('test_scatternd_min_cpu') # min not yet supported
 backend_test.exclude('test_scatternd_max_cpu') # max not yet supported

test/external/external_test_onnx_ops.py

@@ -100,6 +100,25 @@ class TestMainOnnxOps(TestOnnxOps):
     self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=1)
     self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=0)
 
+  def _test_if(self, then_value, else_value):
+    then_out = onnx.helper.make_tensor_value_info("res", onnx.TensorProto.FLOAT, then_value.shape)
+    else_out = onnx.helper.make_tensor_value_info("res", onnx.TensorProto.FLOAT, else_value.shape)
+
+    then_const_node = onnx.helper.make_node("Constant", inputs=[], outputs=["res"], value=onnx.numpy_helper.from_array(then_value))
+    else_const_node = onnx.helper.make_node("Constant", inputs=[], outputs=["res"], value=onnx.numpy_helper.from_array(else_value))
+
+    then_body = onnx.helper.make_graph([then_const_node], "then_body", [], [then_out])
+    else_body = onnx.helper.make_graph([else_const_node], "else_body", [], [else_out])
+
+    self.helper_test_single_op("If", {"cond": np.array(False).astype(bool)}, {"then_branch": then_body, "else_branch": else_body}, ["res"])
+    self.helper_test_single_op("If", {"cond": np.array(True).astype(bool)}, {"then_branch": then_body, "else_branch": else_body}, ["res"])
+
+  def test_if_different_shapes_broadcastable(self):
+    self._test_if(np.array([[1], [2]]).astype(np.float32), np.array([[6, 5, 4, 3, 2, 1]]).astype(np.float32))
+
+  def test_if_different_shapes_not_broadcastable(self):
+    self._test_if(np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32), np.array([[6, 5, 4, 3, 2, 1]]).astype(np.float32))
+
   def test_resize_downsample_scales_linear_align_corners(self):
     # https://github.com/onnx/onnx/blob/main/docs/Operators.md#examples-131
     X = np.array([[[[1, 2, 3, 4], [5, 6, 7, 8]]]], dtype=np.float32)

test/external/fuzz_fast_idiv.py

@@ -1,8 +1,8 @@
 import random
 import z3
 from tinygrad import dtypes
-from tinygrad.uop.spec import z3_renderer, z3_cdiv
-from tinygrad.uop.ops import UOp, graph_rewrite
+from tinygrad.uop.spec import uops_to_z3, z3_cdiv
+from tinygrad.uop.ops import UOp
 from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)
 
@@ -19,8 +19,7 @@ if __name__ == "__main__":
     if expr is None: continue
 
     solver = z3.Solver()
-    z3_sink = graph_rewrite(expr.sink(u), z3_renderer, ctx=(solver, {}))
-    z3_expr, x = z3_sink.src[0].arg, z3_sink.src[1].arg
+    z3_expr, x =uops_to_z3(solver, expr, u)
 
     if solver.check(z3_expr != z3_cdiv(x, d)) == z3.sat:
       assert False, f"Failed: {expr.render()} != x//{d} at x={solver.model()}\nx={u}\nd={d}\n{z3_expr=}\n{x/d=}"

test/external/fuzz_symbolic.py

@@ -1,8 +1,8 @@
 import random, operator
 import z3
 from tinygrad import Variable, dtypes
-from tinygrad.uop.ops import UOp, graph_rewrite
-from tinygrad.uop.spec import z3_renderer
+from tinygrad.uop.ops import UOp
+from tinygrad.uop.spec import uops_to_z3
 from tinygrad.helpers import DEBUG, Context
 
 seed = random.randint(0, 100)
@@ -57,8 +57,7 @@ if __name__ == "__main__":
 
     solver = z3.Solver()
     solver.set(timeout=5000)  # some expressions take very long verify, but its very unlikely they actually return sat
-    z3_sink = graph_rewrite(expr.sink(simplified_expr, u1, u2, u3), z3_renderer, ctx=(solver, {}))
-    z3_expr, z3_simplified_expr = z3_sink.src[0].arg, z3_sink.src[1].arg
+    z3_expr, z3_simplified_expr, v1, v2, v3 = uops_to_z3(solver, expr, simplified_expr, u1, u2, u3)
     check = solver.check(z3_simplified_expr != z3_expr)
     if check == z3.unknown and DEBUG>=1:
       skipped += 1
@@ -69,7 +68,6 @@ if __name__ == "__main__":
             f"expr = {expr.render(simplify=False)}\n")
     elif check == z3.sat:
       m = solver.model()
-      v1, v2, v3 = z3_sink.src[2].arg, z3_sink.src[3].arg, z3_sink.src[4].arg
       n1, n2, n3 = m[v1], m[v2], m[v3]
       u1_val, u2_val, u3_val = u1.const_like(n1.as_long()), u2.const_like(n2.as_long()), u3.const_like(n3.as_long())
       with Context(CORRECT_DIVMOD_FOLDING=1):

test/mockgpu/amd/amdgpu.py

@@ -1,5 +1,6 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
+from test.mockgpu.helpers import _try_dlopen_remu
 from tinygrad.helpers import getbits, to_mv, init_c_struct_t
 import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
 
@@ -24,19 +25,6 @@ WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
 WAIT_REG_MEM_FUNCTION_GEQ = 5 # >=
 
-REMU_PATHS = ["extra/remu/target/release/libremu.so", "libremu.so", "/usr/local/lib/libremu.so",
-              "extra/remu/target/release/libremu.dylib", "libremu.dylib", "/usr/local/lib/libremu.dylib", "/opt/homebrew/lib/libremu.dylib"]
-def _try_dlopen_remu():
-  for path in REMU_PATHS:
-    try:
-      remu = ctypes.CDLL(path)
-      remu.run_asm.restype = ctypes.c_int32
-      remu.run_asm.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32,
-        ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_void_p]
-    except OSError: pass
-    else: return remu
-  print("Could not find libremu.so")
-  return None
 remu = _try_dlopen_remu()
 
 def create_sdma_packets():

test/mockgpu/cuda/cuda.py

@@ -2,16 +2,14 @@ from __future__ import annotations
 from typing import Any
 import ctypes, time
 from tinygrad.runtime.autogen import cuda as orig_cuda
+from test.mockgpu.helpers import _try_dlopen_gpuocelot
 from tinygrad.helpers import mv_address
 
 for attr in dir(orig_cuda):
   if not attr.startswith('__'):
     globals()[attr] = getattr(orig_cuda, attr)
 
-try:
-  gpuocelot_lib = ctypes.CDLL(ctypes.util.find_library("gpuocelot"))
-  gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]  # noqa: E501
-except Exception: pass
+gpuocelot_lib = _try_dlopen_gpuocelot()
 
 # Global state
 class CUDAState:

test/mockgpu/helpers.py

@@ -0,0 +1,29 @@
+import ctypes, ctypes.util
+
+def _try_dlopen_gpuocelot():
+  GPUOCELOT_PATHS = [ctypes.util.find_library("gpuocelot")] if ctypes.util.find_library("gpuocelot") is not None else []
+  GPUOCELOT_PATHS += ["libgpuocelot.so", "/usr/local/lib/libgpuocelot.so",
+                      "libgpuocelot.dylib", "/usr/local/lib/libgpuocelot.dylib", "/opt/homebrew/lib/libgpuocelot.dylib"]
+  for path in GPUOCELOT_PATHS:
+    try:
+      gpuocelot_lib = ctypes.CDLL(path)
+      gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int,
+        ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]
+    except OSError: pass
+    else: return gpuocelot_lib
+  print("Could not find libgpuocelot.so")
+  return None
+
+def _try_dlopen_remu():
+  REMU_PATHS = ["extra/remu/target/release/libremu.so", "libremu.so", "/usr/local/lib/libremu.so",
+               "extra/remu/target/release/libremu.dylib", "libremu.dylib", "/usr/local/lib/libremu.dylib", "/opt/homebrew/lib/libremu.dylib"]
+  for path in REMU_PATHS:
+    try:
+      remu = ctypes.CDLL(path)
+      remu.run_asm.restype = ctypes.c_int32
+      remu.run_asm.argtypes = [ctypes.c_void_p, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32,
+        ctypes.c_uint32, ctypes.c_uint32, ctypes.c_uint32, ctypes.c_void_p]
+    except OSError: pass
+    else: return remu
+  print("Could not find libremu.so")
+  return None

test/mockgpu/nv/nvgpu.py

@@ -2,6 +2,7 @@ import ctypes, ctypes.util, time
 import tinygrad.runtime.autogen.nv_gpu as nv_gpu
 from enum import Enum, auto
 from test.mockgpu.gpu import VirtGPU
+from test.mockgpu.helpers import _try_dlopen_gpuocelot
 from tinygrad.helpers import to_mv, init_c_struct_t
 
 def make_qmd_struct_type():
@@ -16,10 +17,7 @@ def make_qmd_struct_type():
 qmd_struct_t = make_qmd_struct_type()
 assert ctypes.sizeof(qmd_struct_t) == 0x40 * 4
 
-try:
-  gpuocelot_lib = ctypes.CDLL(ctypes.util.find_library("gpuocelot"))
-  gpuocelot_lib.ptx_run.argtypes = [ctypes.c_char_p, ctypes.c_int, ctypes.POINTER(ctypes.c_void_p), ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int, ctypes.c_int]  # noqa: E501
-except Exception: pass
+gpuocelot_lib = _try_dlopen_gpuocelot()
 
 class SchedResult(Enum): CONT = auto(); YIELD = auto() # noqa: E702
 

test/test_const_folding.py

@@ -1,11 +1,10 @@
 import unittest, itertools, math
-from typing import Any
 from tinygrad import Tensor, Device, dtypes
-from tinygrad.dtype import DType
+from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.codegen import full_rewrite_to_sink
-import numpy as np
 from tinygrad.device import is_dtype_supported
+import numpy as np
 from test.helpers import not_support_multi_device
 
 def _check_ast_count(desired_count:int, t:Tensor):
@@ -25,7 +24,7 @@ class TestUnaryOpsConstFolding(unittest.TestCase):
     _check_ast_count(0, Tensor.ones(4).cast(dtypes.int16))
     _check_ast_count(0, Tensor.full(4, fill_value=-1).cast(dtypes.uint16))
 
-  @unittest.expectedFailure  # no two level fold at lazybuffer
+  @unittest.expectedFailure  # no two level fold
   def test_neg_folding(self):
     _check_ast_count(0, Tensor([1, 2, 3]).mul(-1).neg())
     _check_ast_count(0, Tensor([1, 2, 3]).neg().mul(-1))
@@ -104,7 +103,7 @@ class TestBinaryOpsConstFolding(unittest.TestCase):
 
 class TestBitcastConstFolding(unittest.TestCase):
   def test_scalar_bitcast(self):
-    def t(cases: dict[DType, Any]):
+    def t(cases: dict[DType, ConstType]):
       for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
         if not math.isnan(from_v):
           r = full_rewrite_to_sink(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
@@ -165,7 +164,6 @@ class TestMovedConstFolding(unittest.TestCase):
     _check_ast_count(1, Tensor([1.0, 2, 3, 4]) * Tensor.ones(2).pad(((1, 1),)))
 
   def test_cast_padded(self):
-    # NOTE: this is folded due to CAST_BEFORE_VIEW
     if is_dtype_supported(dtypes.int16):
       _check_ast_count(0, Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16))
       np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).cast(dtypes.int16).numpy(), [0, 1, 1, 1, 1, 0])

test/test_dtype.py

@@ -4,9 +4,8 @@ import torch
 from typing import Any, List
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, DEBUG, CI
-from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8
+from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype
 from tinygrad import Device, Tensor, dtypes
-from tinygrad.tensor import _to_np_dtype
 from hypothesis import assume, given, settings, strategies as strat
 from test.helpers import rand_for_dtype
 from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX
@@ -24,6 +23,10 @@ def get_available_cast_dtypes(dtype: DType) -> List[DType]:
   # dont cast internal dtypes
   return [v for k, v in DTYPES_DICT.items() if v != dtype and is_dtype_supported(v) and not k.startswith("_")]
 
+def _to_torch_storage_type(dtype:DType):
+  if dtype == dtypes.bfloat16: return torch.float32
+  return _to_torch_dtype(dtype)
+
 def _test_to_np(a:Tensor, np_dtype, target):
   if DEBUG >= 2: print(a)
   na = a.numpy()
@@ -46,10 +49,10 @@ def _test_cast(a:Tensor, target_dtype:DType):
 
   _test_op(lambda: a.cast(target_dtype), target_dtype, list(a.numpy().astype(_to_np_dtype(target_dtype))))
 def _test_bitcast(a:Tensor, target_dtype:DType, target=None):
-  if target_dtype == dtypes.bfloat16: raise unittest.SkipTest("no test for bf16 bitcast yet")
   if getenv("PTX") and a.dtype == dtypes.int8 and target_dtype.itemsize != a.dtype.itemsize:
     raise unittest.SkipTest("shape changing bitcast of int8 broken on PTX")
-  _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or a.numpy().view(_to_np_dtype(target_dtype)).tolist())
+  expected = torch.tensor(a.tolist(), dtype=_to_torch_storage_type(a.dtype)).view(_to_torch_dtype(target_dtype))
+  _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or expected.tolist())
 
 class TestDType(unittest.TestCase):
   DTYPE: Any = None
@@ -126,7 +129,7 @@ class TestDType(unittest.TestCase):
 
   def test_finfo(self):
     if self.DTYPE not in [dtypes.float16, dtypes.bfloat16, dtypes.float32, dtypes.float64]: return
-    info = np.finfo(_to_np_dtype(self.DTYPE))
+    info = ml_dtypes.finfo(ml_dtypes.bfloat16 if self.DTYPE is dtypes.bfloat16 else _to_np_dtype(self.DTYPE))
     assert info.bits == self.DTYPE.itemsize*8
     assert info.nexp == dtypes.finfo(self.DTYPE)[0]
     assert info.nmant == dtypes.finfo(self.DTYPE)[1]
@@ -299,10 +302,10 @@ class TestBitCast(unittest.TestCase):
   @given(strat.sampled_from(dtype_ints + dtype_floats), strat.sampled_from(dtype_ints + dtype_floats))
   def test_shape_change_bitcast(self, dt1, dt2):
     # NOTE: this has to be assume to prevent hypothesis from skipping all samples
-    assume(dt2 != dtypes.bfloat16 and dt1 != dtypes.bfloat16) # no test for bf16 bitcast yet
     assume(not (getenv("PTX") and dt1 == dtypes.int8)) # TODO: bitcasting int8 fails in PTX
     data = rand_for_dtype(dt1, 32).reshape(2, 2, 8)
-    _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, data.view(_to_np_dtype(dt2)).tolist())
+    expected = torch.tensor(data.tolist(), dtype=_to_torch_storage_type(dt1)).view(_to_torch_dtype(dt2))
+    _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, expected.tolist())
 
   def test_shape_change_bitcast_exceptions(self):
     with self.assertRaises(RuntimeError):
@@ -342,6 +345,9 @@ class TestUint64DType(TestDType):
 
 class TestBoolDType(TestDType): DTYPE = dtypes.bool
 
+@unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), f"no bfloat16 on {Device.DEFAULT}")
+class TestBFloat16Type(TestDType): DTYPE = dtypes.bfloat16
+
 class TestPtrDType(unittest.TestCase):
   def test_vec_double(self):
     dt1 = dtypes.float.vec(4).ptr().vec(4)
@@ -418,7 +424,7 @@ class TestDtypeUsage(unittest.TestCase):
 class TestOpsBFloat16(unittest.TestCase):
   def test_cast(self):
     # TODO: helper_test_op breaks in unrelated part
-    # TODO: wrong output with GPU=1 / PYTHON=1 on mac
+    # TODO: wrong output with GPU=1 on mac
     data = [60000.0, 70000.0, 80000.0]
     np.testing.assert_allclose(Tensor(data).cast("bfloat16").numpy(), torch.tensor(data).type(torch.bfloat16).float().numpy())
 

test/test_linearizer.py

@@ -12,7 +12,7 @@ from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, lower_schedule, CompiledRunner, get_program
 from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import prod, Context, getenv, CI, flatten, dedup, AMX, AMD_LLVM
-from tinygrad.dtype import DType, dtypes, AddrSpace
+from tinygrad.dtype import DType, dtypes, PtrDType, AddrSpace
 from tinygrad.codegen import apply_rewrites, rewrites_for_views
 
 def push_views(ast): return apply_rewrites(ast, rewrites_for_views)
@@ -33,11 +33,10 @@ def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axi
   r = a.matmul(b, dtype=dtype_out)
   if dtype_in == dtypes.bfloat16: r = r.float()
   realized_ast, bufs = helper_realized_ast(r)
-  k = Kernel(realized_ast)
-  k.apply_tensor_cores(use_tensor_cores, axis=axis, tc_select=tc_select, tc_opt=tc_opt)
-  prg = CompiledRunner(replace(get_program(k.get_optimized_ast(), k.opts), device=Device.DEFAULT))
+  opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
+  prg = CompiledRunner(replace(get_program(realized_ast, opts=opts), device=Device.DEFAULT))
   if use_tensor_cores == 1: assert len([uop for uop in prg.p.uops if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
-  assert len([x for x in k.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
+  assert len([x for x in prg.p.uops[-1].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
   prg.exec(bufs)
   if dtype_in == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
   elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = 1e-2, 1e-2
@@ -134,7 +133,7 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
-    # RANGE -> LOAD -> RANGE -> ASSIGN
+    # RANGE -> LOAD -> RANGE -> STORE
     #assert any(x.op is Ops.LOAD for x in uops[ranges[0]:ranges[1]])
 
   def test_three_nested_range(self):
@@ -144,7 +143,7 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
-    # RANGE -> RANGE -> LOAD -> RANGE -> ASSIGN
+    # RANGE -> RANGE -> LOAD -> RANGE -> STORE
     # NOTE: nothing should toposort between the first two ranges
     #assert ranges[0]+1 == ranges[1]
     #assert any(x.op is Ops.LOAD for x in uops[ranges[1]:ranges[2]])
@@ -155,7 +154,7 @@ class TestLinearizer(unittest.TestCase):
     lin = helper_linearizer_opt(out, wanna_output=[24])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
-    # RANGE -> ALU -> RANGE -> ALU + LOAD -> ASSIGN
+    # RANGE -> ALU -> RANGE -> ALU + LOAD -> STORE
     assert any(x.op in GroupOp.ALU for x in uops[ranges[0]:ranges[1]])
     assert not any(x.op is Ops.LOAD for x in uops[ranges[0]:ranges[1]])
     assert any(x.op in {*GroupOp.ALU, Ops.LOAD} for x in uops[ranges[1]:])
@@ -167,7 +166,7 @@ class TestLinearizer(unittest.TestCase):
     lin = helper_linearizer_opt(out, wanna_output=[(a.numpy()+b.numpy()[0]).sum()+b.numpy()])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
-    # LOAD -> RANGE -> LOAD -> ASSIGN
+    # LOAD -> RANGE -> LOAD -> STORE
     assert len([x for x in uops[:ranges[0]] if x.op is Ops.LOAD]) == 1
 
   def test_range_outer_op_before_phi_nested_range(self):
@@ -179,11 +178,11 @@ class TestLinearizer(unittest.TestCase):
     ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
     assert len(ranges) == 1 # NOTE: it collapses now
     #if getenv("PTX"):
-    # LOAD -> RANGE -> CAST -> ALU -> ALU -> LOAD -> ALU -> RANGE -> ALU -> ASSIGN
+    # LOAD -> RANGE -> CAST -> ALU -> ALU -> LOAD -> ALU -> RANGE -> ALU -> STORE
     #  assert uops[ranges[0]-2].op is Ops.LOAD
     #  assert ranges[1] == ranges[0]+6
     #  assert [x.op for x in uops[ranges[1]-2:ranges[1]]] == [Ops.LOAD, Ops.ALU]
-    # LOAD -> RANGE -> LOAD -> ALU -> RANGE -> ASSIGN
+    # LOAD -> RANGE -> LOAD -> ALU -> RANGE -> STORE
     #else:
     #  assert uops[ranges[0]-2].op is Ops.LOAD
     #  assert ranges[1] == ranges[0]+3
@@ -195,7 +194,7 @@ class TestLinearizer(unittest.TestCase):
     out = a.sum() * a.sum()
     lin = helper_linearizer_opt(out, wanna_output=[a.numpy().sum()*a.numpy().sum()])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
-    # RANGE -> LOAD -> ASSIGN -> ALU
+    # RANGE -> LOAD -> STORE -> ALU
     end = max(i for i,u in enumerate(uops) if u.op is Ops.ENDRANGE)
     # the INDEX can be first
     assert uops[end+1].op in GroupOp.ALU or uops[end+2].op in GroupOp.ALU
@@ -206,7 +205,7 @@ class TestLinearizer(unittest.TestCase):
     out = a.reshape(2, 1).expand(2, 3).sum() + a.reshape(2, 1).expand(2, 3).sum()
     lin = helper_linearizer_opt(out, wanna_output=[(np.broadcast_to(a.numpy().reshape(2, 1), (2, 3))).sum()*2])[0]
     uops = get_program(lin.get_optimized_ast(), lin.opts).uops
-    # RANGE -> LOAD -> ASSIGN -> ALU
+    # RANGE -> LOAD -> STORE -> ALU
     end = max(i for i,u in enumerate(uops) if u.op is Ops.ENDRANGE)
     # the INDEX can be first
     assert uops[end+1].op in GroupOp.ALU or uops[end+2].op in GroupOp.ALU
@@ -328,11 +327,7 @@ class TestLinearizer(unittest.TestCase):
       n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
       a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
       r = a.matmul(b, dtype=tc.dtype_out)
-      sched = r.schedule()
-      realized_ast = push_views(sched[-1].ast)
-      kernel = Kernel(realized_ast)
-      kernel.apply_tensor_cores(1, axis=0, tc_select=-1, tc_opt=2)
-      prg = get_program(kernel.get_optimized_ast(), kernel.opts)
+      prg = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
       if Device.DEFAULT == "LLVM":
         assert "0x201000" in prg.src
       elif Device.DEFAULT == "AMD" and AMD_LLVM:
@@ -353,7 +348,7 @@ class TestLinearizer(unittest.TestCase):
   # Internal bug: zero-stride dimensions combined with a mask may produce wrong index/valid for pad == 1 on AMD
   @unittest.skipUnless((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and getenv("EMULATE_AMD")), "test for AMD's tc")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
-  @unittest.expectedFailure
+  @unittest.skip("warp elements not duplicated properly across lanes")
   def test_tensor_cores_padded_amd(self):
     for tc in Device[Device.DEFAULT].renderer.tensor_cores:
       if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
@@ -423,9 +418,9 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "CPU does not support using a different type for accumulation")
@@ -434,37 +429,43 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "CPU does not support using a different type for accumulation")
   def test_tensor_cores_unroll_casted_phi_with_children(self):
-    # all ASSIGN children are outside the loop
+    # all STORE children are outside the loop
     tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
     x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out).relu()
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4)]], apply_tc=True, atol=3e-2, rtol=1e-3)[-1]
-    for u in get_program(k.get_optimized_ast(), k.opts).uops:
+    for u in get_program(k.ast, k.opts, k.applied_opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
-        assert u.src[-1].src[0].op != Ops.ASSIGN
+        assert u.src[-1].src[0].op != Ops.STORE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
   def test_simple_unroll_no_between_phi_dependencies(self):
     x, y = Tensor.rand(128, 128), Tensor.rand(128, 128)
     r = (x@y).relu()
     k = helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4)]])[-1]
-    # the uops graph is RANGE -> DEFINE_ACC -> 4x ALU -> 4x ASSIGN -> ENDRANGE
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    # the uops graph is DEFINE_REG -> 4x STORE 0.0 -> RANGE -> 4x ALU -> 4x STORE -> ENDRANGE
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
+    begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
+    end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
+    for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
     for u in uops:
-      if u.op is Ops.ASSIGN:
-        assert u.src[1].op in GroupOp.ALU
-      # children of ASSIGN are placed after ENDRANGE
-      if any(x.op is Ops.ASSIGN for x in u.src):
-        end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
+      if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace is AddrSpace.REG:
+        if uops.index(u) < begin_range:
+          assert u.src[1].op is Ops.CONST
+        else:
+          assert u.src[1].op in GroupOp.ALU
+          assert begin_range < uops.index(u) < end_range
+      # children of STORE are placed after ENDRANGE
+      if any(x.op is Ops.STORE and x.src[1].op in GroupOp.ALU for x in u.src):
         assert end_range < uops.index(u)
 
   def test_grouped_dims(self):
@@ -543,7 +544,7 @@ class TestLinearizer(unittest.TestCase):
     # shrink so that the dims do not collapse
     t = Tensor.ones(5, 6, 7).contiguous().realize().shrink(((0, 4), (0, 5), (0, 6)))
     k = helper_linearizer_opt(t+1)[0]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     idxs = dedup([uop for uop in uops if uop.op is Ops.SPECIAL])
     idxs = sorted(idxs, key=lambda uop: uop.arg[0])
     assert idxs[0].arg == ('gidx0', 6), idxs[0].arg
@@ -583,12 +584,13 @@ class TestLinearizer(unittest.TestCase):
   def test_phi_simplification(self):
     def helper(t, max_ops=0):
       k = helper_linearizer_opt(t)[-1]
-      uops = get_program(k.get_optimized_ast(), k.opts).uops
+      uops = get_program(k.ast, k.opts, k.applied_opts).uops
       # ignore kernel optimized IF statements for now
       if if_op:=next((u for u in uops if u.op is Ops.IF), None):
         uops = uops[:uops.index(if_op)]
       assert len(set([u.op for u in uops if u.op in {Ops.RANGE, Ops.SPECIAL}])) == 1, "has either specials or ranges, not both"
-      assert len([u for u in uops if u.op is Ops.ASSIGN]) == 0, "ASSIGN should have been simplified"
+      reg_stores = [u for u in uops if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace == AddrSpace.REG]
+      assert len(reg_stores) == 0, "STORE to reg should have been simplified"
       # TODO: once uops track min/max this will be fixed
       #assert len([u for u in uops if u.op is Ops.MAX]) <= max_ops, "no unnecessary MAX ops"
 
@@ -614,7 +616,7 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.randn(64,64), Tensor.randn(64,64)
     out = x.matmul(y)
     k = helper_linearizer_opt(out)[-1]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     # check that the float4 cast collapses
     store_vals = [u.src[1] for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
     for val in store_vals:
@@ -639,7 +641,7 @@ class TestLinearizer(unittest.TestCase):
     x = Tensor.randn((4,3,6,6)).realize()
     out = x.flip((0,1)).contiguous()
     k = helper_linearizer_opt(out)[-1]
-    store_val = [u.src[1] for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    store_val = [u.src[1] for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert store_val.dtype == dtypes.float.vec(4) and store_val.op is not Ops.VECTORIZE
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -652,7 +654,7 @@ class TestLinearizer(unittest.TestCase):
             Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 2)] # upcast accs in both reduces
     k = helper_linearizer_opt(out, opts=[opt])[-1]
     def get_recursive(uop): return set.union(set(uop.src), [uop], *[get_recursive(v) for v in uop.src])
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     local_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_LOCAL for x in get_recursive(u.src[0]))]
     global_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_GLOBAL for x in get_recursive(u.src[0]))]
     barrier = [u for u in uops if u.op is Ops.BARRIER][0]
@@ -672,7 +674,7 @@ class TestLinearizer(unittest.TestCase):
     x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
     r = (x@y).relu()
     k = helper_linearizer_opt(r)[-1]
-    uops = get_program(k.get_optimized_ast(), k.opts).uops
+    uops = get_program(k.ast, k.opts, k.applied_opts).uops
     stores = [u for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
 
     # the float4 value stores directly in lds and we skip upcast
@@ -698,7 +700,7 @@ class TestLinearizer(unittest.TestCase):
         Opt(op=OptOps.LOCAL, axis=1, arg=2), Opt(op=OptOps.UPCAST, axis=3, arg=2)
     ]
     k = helper_linearizer_ast(ast, [Tensor.randn(240*40).realize()], opts=[opt])[-1]
-    out = [u for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    out = [u for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert out.src[1].op is Ops.VECTORIZE and out.src[1].dtype == dtypes.float.vec(4)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -716,7 +718,7 @@ class TestLinearizer(unittest.TestCase):
             Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8),
             Opt(op=OptOps.UPCAST, axis=1, arg=0), Opt(op=OptOps.UPCAST, axis=0, arg=2)]
     k = helper_linearizer_ast(ast, [Tensor.randn(8*32).realize()], opts=[opt])[-1]
-    out = [u for u in get_program(k.get_optimized_ast(), k.opts).uops if u.op is Ops.STORE][0]
+    out = [u for u in get_program(k.ast, k.opts, k.applied_opts).uops if u.op is Ops.STORE][0]
     assert out.src[1].op is Ops.VECTORIZE and out.src[1].dtype.count != 1
 
 @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need backends that support float4")
@@ -1047,7 +1049,7 @@ def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[]
   outbufs = [real_bufs[x.src[0].base.arg] for x in realized_ast.src]
   device = real_bufs[0].device
 
-  def get_prg(k:Kernel): return CompiledRunner(replace(get_program(k.get_optimized_ast(), k.opts), device=device))
+  def get_prg(k:Kernel): return CompiledRunner(replace(get_program(k.ast, k.opts, k.applied_opts), device=device))
 
   def check_opt(opts, create_k, expected_color_size):
     k = create_k()

test/test_linearizer_dumb.py

@@ -9,7 +9,6 @@ from tinygrad.uop.ops import UOp, Ops
 from tinygrad.helpers import getenv
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.codegen.opt.search import Opt, OptOps
-from tinygrad.codegen.opt.kernel import Kernel
 from tinygrad.engine.realize import get_program
 
 class TestLinearizerDumb(unittest.TestCase):
@@ -36,9 +35,7 @@ class TestLinearizerDumb(unittest.TestCase):
           UOp(Ops.CONST, dtypes.half, arg=0.0, src=(
              x16,)),)),)),))
     opts = [Opt(op=OptOps.TC, axis=2, arg=(-1, 2, 1)), Opt(op=OptOps.UPCAST, axis=2, arg=0), Opt(op=OptOps.UNROLL, axis=1, arg=0)]
-    k = Kernel(ast, opts=Device["METAL"].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device["METAL"].renderer, opts)
     print(prg.src)
     Device[Device.DEFAULT].compiler.compile_cached(prg.src)
     gate_count = len([x for x in prg.src.splitlines() if "if" in x])
@@ -75,9 +72,7 @@ class TestLinearizerDumb(unittest.TestCase):
             UOp(Ops.CONST, dtypes.int, arg=1000, src=(
                x14,)),)),)),)),))
     opts = [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
     assert prg.uops is not None and not any(uop.op is Ops.MAX for uop in prg.uops), "leftover MAX"
 
@@ -93,9 +88,7 @@ class TestLinearizerDumb(unittest.TestCase):
             UOp(Ops.VIEW, dtypes.float.ptr(25), arg=ShapeTracker(views=(View(shape=(26, 49), strides=(0, -1), offset=48, mask=((0, 26), (24, 49)), contiguous=False), View(shape=(25, 25), strides=(1, 50), offset=0, mask=None, contiguous=False))), src=(
               UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(25), arg=1, src=()),)),)),)),)),))
     opts = [Opt(op=OptOps.GROUP, axis=0, arg=0), Opt(op=OptOps.PADTO, axis=0, arg=32), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
     if_uops = [u for u in prg.uops if u.op is Ops.IF]
     self.assertIn(len(if_uops), {1,2,3})
@@ -135,8 +128,7 @@ class TestLinearizerDumb(unittest.TestCase):
                 UOp(Ops.LOAD, dtypes.half, arg=None, src=(
                   UOp(Ops.VIEW, dtypes.half.ptr(131072000), arg=ShapeTracker(views=(View(shape=(4096, 32000, 1), strides=(1, 4096, 0), offset=0, mask=None, contiguous=False),)), src=(
                     UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(131072000), arg=2, src=()),)),)),)),)),)),)),)),))
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer)
     print(prg.src)
 
   @unittest.expectedFailure
@@ -163,11 +155,9 @@ class TestLinearizerDumb(unittest.TestCase):
               UOp(Ops.VIEW, dtypes.float.ptr(18), arg=ShapeTracker(views=(View(shape=(3, 6), strides=(6, 1), offset=0, mask=None, contiguous=True),)), src=(
                 UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(18), arg=2, src=()),)),)),)),)),)),))
     opts = [Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
-    load_idxs = [x.src[1] for x in k.uops if x.op is Ops.LOAD and x.src[0].arg == 2]
+    load_idxs = [x.src[1] for x in prg.uops if x.op is Ops.LOAD and x.src[0].arg == 2]
     assert load_idxs[0] < load_idxs[1], f"first loaded idx {load_idxs[0].arg} then {load_idxs[1].arg}!"
 
   @unittest.expectedFailure
@@ -187,11 +177,9 @@ class TestLinearizerDumb(unittest.TestCase):
               UOp(Ops.VIEW, dtypes.float.ptr(1040), arg=ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 4, 1, 4, 3, 3), strides=(260, 13, 1, 0, 0, 0, 65, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=(
                 UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1040), arg=2, src=()),)),)),)),)),)),))
     opts = [Opt(op=OptOps.UPCAST, axis=3, arg=0), Opt(op=OptOps.UPCAST, axis=2, arg=0)]
-    k = Kernel(ast, opts=Device[Device.DEFAULT].renderer)
-    k.apply_opts(opts)
-    prg = get_program(k.get_optimized_ast(), k.opts)
+    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
     print(prg.src)
-    store_idxs = [x.src[1] for x in k.uops if x.op is Ops.STORE]
+    store_idxs = [x.src[1] for x in prg.uops if x.op is Ops.STORE]
     for i in range(len(store_idxs) - 1):
       first_bounds = store_idxs[i].vmin+store_idxs[i].vmax
       next_bounds = store_idxs[i+1].vmin+store_idxs[i+1].vmax

test/test_memory_planner.py

@@ -120,5 +120,19 @@ class TestMemoryPlanner(unittest.TestCase):
     ]
     check_assign(bs)
 
+  def test_very_small_buffers(self):
+    bs = [
+      [b(0, pin=True), b(1, size=32)],
+      [b(3, size=4), b(4, size=6)],
+    ]
+    check_assign(bs)
+
+  def test_very_big_buffers(self):
+    bs = [
+      [b(0, pin=True), b(1, size=34359738368000)],
+      [b(3, size=1 << 128), b(4, size=1 << 64)],
+    ]
+    check_assign(bs)
+
 if __name__ == "__main__":
   unittest.main()

test/test_outerworld_range.py

@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, nn, Variable, UOp, dtypes
+from tinygrad import Tensor, nn, Variable, UOp
 
 # outerworld range should support three things
 # 1. full optimizer steps (test_model_bound_range)
@@ -136,7 +136,7 @@ class TestOuterworldRange(unittest.TestCase):
   def test_model_bound_range(self):
     m, opt = get_model_and_opt()
     # TODO: should ranges be unique so you don't have to pass in the -1?
-    rng = UOp.range(dtypes.int, self.STEPS, -1)
+    rng = UOp.range(self.STEPS, -1)
     vib = Variable('i', 0, self.STEPS-1).bind(rng)
     loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
     loss.backward()

test/test_rangeify.py

@@ -1,6 +1,7 @@
 import unittest
 from tinygrad import Tensor
-from tinygrad.helpers import RANGEIFY
+from tinygrad.helpers import RANGEIFY, Context, GlobalCounters
+from tinygrad.uop.ops import UOp
 
 N = 256
 
@@ -11,6 +12,26 @@ class TestRangeify(unittest.TestCase):
     ba = A.expand(N, N)
     ((ba+1).sum(axis=1) + (ba+2).sum(axis=0)).realize()
 
+  def test_partial_contig(self):
+    A = Tensor.empty(64, 64, 64)
+    ret = A.sum(axis=2).contiguous(arg=(1,)).sum(axis=1)
+    ret.realize()
+
+  def test_double_gemm_real(self):
+    def go():
+      with Context(DEBUG=0):
+        Tensor.manual_seed(1337)
+        A,B,C = [Tensor.randn(N, N) for _ in range(3)]
+        Tensor.realize(A, B, C)
+      GlobalCounters.reset()
+      return (A@B@C).realize()
+    rng = go()
+    with Context(RANGEIFY=0, DEBUG=2):
+      ref = go()
+      mse = ((rng-ref)**2).sum().item()
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-2)
+
   def test_double_gemm(self):
     A = Tensor.empty(N, N)
     B = Tensor.empty(N, N)
@@ -96,17 +117,30 @@ class TestRangeify(unittest.TestCase):
     out.realize()
 
   def test_flash_attention(self):
-    BS = 4
-    HEADS = 2
-    MATDIM = 16
-    EMB = 8
-    q = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    k = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    v = Tensor.empty(BS, HEADS, MATDIM, EMB)
-    q.scaled_dot_product_attention(k, v).realize()
+    BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
 
-from tinygrad import dtypes
-from tinygrad.uop.ops import UOp
+    # bigger
+    #BS, HEADS, SEQLEN, EMB = 4, 32, 1024, 64
+
+    # llama 8B
+    #BS, HEADS, SEQLEN, EMB = 4, 32, 2048, 128
+
+    def fa():
+      Tensor.manual_seed(1337)
+      with Context(DEBUG=0): q,k,v = [Tensor.rand(BS, HEADS, SEQLEN, EMB).contiguous().realize() for _ in range(3)]
+      return q.scaled_dot_product_attention(k, v).realize()
+
+    with Context(DEBUG=4):
+      GlobalCounters.reset()
+      ret = fa()
+    with Context(RANGEIFY=0):
+      with Context(DEBUG=2):
+        GlobalCounters.reset()
+        cmp = fa()
+      with Context(DEBUG=0):
+        mse = ((cmp-ret)**2).sum().item()
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-6)
 
 # contiguous + reduce can support ranges?
 
@@ -116,7 +150,7 @@ class TestOuterworld(unittest.TestCase):
     t = Tensor.rand(10, 10).realize()
 
     # passthrough ranges
-    a = UOp.range(dtypes.int, 10, -1)
+    a = UOp.range(10, -1)
     sel = t[a]
     cpy = sel.contiguous(a).realize()
 
@@ -126,7 +160,7 @@ class TestOuterworld(unittest.TestCase):
     t = Tensor.rand(10, 10).realize()
 
     # passthrough ranges
-    a = UOp.range(dtypes.int, 10, -1)
+    a = UOp.range(10, -1)
     sel = t[9-a]
     cpy = sel.contiguous(a).realize()
 
@@ -138,7 +172,7 @@ class TestOuterworld(unittest.TestCase):
     x = Tensor.ones(3, 10, 2).contiguous()
 
     # vmap across axis 0
-    a = UOp.range(dtypes.int, 3, -1)
+    a = UOp.range(3, -1)
     out = f(x[a])
     out = out.contiguous(a)
 
@@ -152,7 +186,7 @@ class TestOuterworld(unittest.TestCase):
 
     manual = (x @ W[0] @ W[1] @ W[2]).contiguous().realize()
 
-    a = UOp.range(dtypes.int, 3, -1)
+    a = UOp.range(3, -1)
     x = x.assign(x @ W[a])
     out = x.contiguous(a)[-1].contiguous().realize()
 

test/test_renderer_failures.py

@@ -25,7 +25,8 @@ def _test_uop_result(inputs:List[Tensor], stores:List[UOp], local_size=None):
       initial_value=np.zeros(sz, dtype=_to_np_dtype(dtype)).data) for u in uops if u.op is Ops.STORE]
   inbufs = [cast(UOp,x.uop).base.buffer for x in inputs]
   src = Device[Device.DEFAULT].renderer.render(uops)
-  ei = CompiledRunner(ProgramSpec("test", src, Device.DEFAULT, uops[-1], uops=uops, local_size=local_size))
+  ei = CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test",
+                                  src, Device.DEFAULT, uops[-1], uops=uops, local_size=local_size))
   ei.exec(outbufs+inbufs)
   return [np.frombuffer(x.as_buffer(), _to_np_dtype(x.dtype)) for x in outbufs]
 

test/test_schedule.py

@@ -1050,6 +1050,14 @@ class TestSchedule(unittest.TestCase):
       compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
       np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
 
+    with Context(FUSE_ATTENTION=1):
+      out = Tensor.scaled_dot_product_attention(q,k,v)
+      run_schedule(check_schedule(out, 1))
+      if getenv("CHECK", 1):
+        import torch
+        compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
+        np.testing.assert_allclose(out.numpy(), compare.numpy(), atol=1e-6, rtol=1e-3)
+
   def test_ugly_reduceop_pairing(self):
     Tensor.manual_seed(0)
     a = Tensor.randn(4, 32).realize()

test/test_symbolic_jit.py

@@ -2,50 +2,41 @@ import unittest
 
 from test.helpers import assert_jit_cache_len
 from tinygrad import Variable, Tensor, TinyJit
-from tinygrad.helpers import Context
 import numpy as np
 
 class TestSymbolicJit(unittest.TestCase):
-  def setUp(self):
-    # A lot of these test are out of bounds, so we ignore the bounds check
-    self.context = Context(IGNORE_OOB=1)
-    self.context.__enter__()
-
-  def tearDown(self):
-    self.context.__exit__(None, None, None)
-
   def test_plus1(self):
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      symbolic = jf(a.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a).numpy()
+      symbolic = jf(a[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_add(self):
     def f(a, b): return (a+b).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, i)
-      symbolic = jf(a.reshape(3, vi), b.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i], b[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_matmul(self):
     def f(a, b): return (a@b).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = jf(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
@@ -55,119 +46,119 @@ class TestSymbolicJit(unittest.TestCase):
       s = (s+s).realize() # this one does not have symbols in input
       return s
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = jf(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 2)
 
   def test_attention(self):
     def f(q, k, v): return Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)).realize()
     jf = TinyJit(f)
+    q = Tensor.rand(2, 1, 4, 8)
+    k = Tensor.rand(2, 10, 4, 8)
+    v = Tensor.rand(2, 10, 4, 8)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      q = Tensor.rand(2, 1, 4, 8)
-      k = Tensor.rand(2, i, 4, 8)
-      v = Tensor.rand(2, i, 4, 8)
-      symbolic = jf(q, k.reshape(2, vi, 4, 8), v.reshape(2, vi, 4, 8)).reshape(2, 4, 1, 8).numpy()
-      expected = f(q, k, v).numpy()
+      symbolic = jf(q, k[:, :vi], v[:, :vi]).reshape(2, 4, 1, 8).numpy()
+      expected = f(q, k[:, :i], v[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 5)
 
   def test_cat_dim0(self):
     def f(a, b): return a.cat(b, dim=0).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(2, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
-      b = Tensor.rand(2, 3)
-      symbolic = jf(a.reshape(vi, 3), b).reshape(i+2, 3).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:vi], b).reshape(i+2, 3).numpy()
+      expected = f(a[:i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim1(self):
     def f(a, b): return a.cat(b, dim=1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 2)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, 2)
-      symbolic = jf(a.reshape(3, vi), b).reshape(3, i+2).numpy()
-      expected = f(a, b).numpy()
+      symbolic = jf(a[:, :vi], b).reshape(3, i+2).numpy()
+      expected = f(a[:, :i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim0_two_vars(self):
     def f(a, b): return a.cat(b, dim=0).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(j, 3)
-        symbolic = jf(a.reshape(vi, 3), b.reshape(vj, 3)).reshape(i+j, 3).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vi], b[:vj]).reshape(i+j, 3).numpy()
+        expected = f(a[:i], b[:j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_cat_dim1_two_vars(self):
     def f(a, b): return a.cat(b, dim=1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(3, i)
-        b = Tensor.rand(3, j)
-        symbolic = jf(a.reshape(3, vi), b.reshape(3, vj)).reshape(3, i+j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:, :vi], b[:, :vj]).reshape(3, i+j).numpy()
+        expected = f(a[:, :i], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_two_vars_plus1_ij(self):
     def f(a, b): return (a@b+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(3, j)
-        symbolic = jf(a.reshape(vi, 3), b.reshape(3, vj)).reshape(i, j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vi, :], b[:, :vj]).reshape(i, j).numpy()
+        expected = f(a[:i, :], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_two_vars_plus1_ji(self):
     def f(a, b): return (a@b+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(j, 3)
-        b = Tensor.rand(3, i)
-        symbolic = jf(a.reshape(vj, 3), b.reshape(3, vi)).reshape(j, i).numpy()
-        expected = f(a, b).numpy()
+        symbolic = jf(a[:vj, :], b[:, :vi]).reshape(j, i).numpy()
+        expected = f(a[:j, :], b[:, :i]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
     assert_jit_cache_len(jf, 1)
 
   def test_jit_symbolic_shape_mismatch(self):
     @TinyJit
     def add(a, b): return (a+b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i).reshape(3, vi)
-      b = Tensor.rand(3, i).reshape(3, vi)
-      add(a, b)
+      add(a[:, :vi], b[:, :vi])
     vi2 = Variable("i", 1, 10).bind(7)
-    a = Tensor.rand(3, 7).reshape(3, vi2)
-    bad = Tensor.rand(4, 7).reshape(4, vi2)
+    a = Tensor.rand(3, 7)[:, :vi2]
+    bad = Tensor.rand(4, 7)[:, :vi2]
     with self.assertRaises(AssertionError):
       add(a, bad)
 
@@ -175,9 +166,9 @@ class TestSymbolicJit(unittest.TestCase):
     # shrink is a movement, so we pair it with a simple function to test the JIT interaction
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
       symbolic = a.shrink(((3,5),(vi,vi+2)))
       symbolic = jf(symbolic).numpy()
       expected = f(a.shrink(((3,5),(i,i+2)))).numpy()
@@ -188,9 +179,9 @@ class TestSymbolicJit(unittest.TestCase):
     # slice is a movement, so we pair it with a simple function to test the JIT interaction
     def f(a): return (a+1).realize()
     jf = TinyJit(f)
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
       symbolic = a[3:5, vi:vi+2]
       symbolic = jf(symbolic).numpy()
       expected = f(a[3:5, i:i+2]).numpy()
@@ -212,11 +203,11 @@ class TestSymbolicJit(unittest.TestCase):
   def test_ones_sum(self):
     def f(a): return a.sum().realize()
     jf = TinyJit(f)
+    t = Tensor.ones(10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      t = Tensor.ones(i)
-      symbolic = jf(t.reshape(vi)).item()
-      expected = f(t).item()
+      symbolic = jf(t[:vi]).item()
+      expected = f(t[:i]).item()
       np.testing.assert_equal(symbolic, expected)
 
   def test_mean(self):
@@ -226,22 +217,22 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
+    c = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      # aixs = None
-      a = Tensor.rand(i, 3)
-      symbolic = jf(a.reshape(vi, 3)).numpy()
-      expected = a.mean().numpy()
+      # axis = None
+      symbolic = jf(a[:vi]).numpy()
+      expected = a[:i].mean().numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 0
-      a = Tensor.rand(i, 3)
-      symbolic = jf0(a.reshape(vi, 3)).numpy()
-      expected = a.mean(0).numpy()
+      # axis = 0
+      symbolic = jf0(b[:vi]).numpy()
+      expected = b[:i].mean(0).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 1
-      a = Tensor.rand(i, 3)
-      symbolic = jf1(a.reshape(vi, 3)).reshape(i).numpy()
-      expected = a.mean(1).numpy()
+      # axis = 1
+      symbolic = jf1(c[:vi]).reshape(i).numpy()
+      expected = c[:i].mean(1).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_mean_2d(self):
@@ -251,24 +242,24 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 10)
+    b = Tensor.rand(10, 10)
+    c = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        # aixs = None
-        a = Tensor.rand(i, j)
-        symbolic = jf(a.reshape(vi, vj)).numpy()
-        expected = a.mean().numpy()
+        # axis = None
+        symbolic = jf(a[:vi, :vj]).numpy()
+        expected = a[:i, :j].mean().numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 0
-        a = Tensor.rand(i, j)
-        symbolic = jf0(a.reshape(vi, vj)).reshape(j).numpy()
-        expected = a.mean(0).numpy()
+        # axis = 0
+        symbolic = jf0(b[:vi, :vj]).reshape(j).numpy()
+        expected = b[:i, :j].mean(0).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 1
-        a = Tensor.rand(i, j)
-        symbolic = jf1(a.reshape(vi, vj)).reshape(i).numpy()
-        expected = a.mean(1).numpy()
+        # axis = 1
+        symbolic = jf1(c[:vi, :vj]).reshape(i).numpy()
+        expected = c[:i, :j].mean(1).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var(self):
@@ -278,22 +269,22 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
+    c = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      # aixs = None
-      a = Tensor.rand(i, 3)
-      symbolic = jf(a.reshape(vi, 3)).numpy()
-      expected = a.var().numpy()
+      # axis = None
+      symbolic = jf(a[:vi]).numpy()
+      expected = a[:i].var().numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 0
-      a = Tensor.rand(i, 3)
-      symbolic = jf0(a.reshape(vi, 3)).numpy()
-      expected = a.var(0).numpy()
+      # axis = 0
+      symbolic = jf0(b[:vi]).numpy()
+      expected = b[:i].var(0).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-      # aixs = 1
-      a = Tensor.rand(i, 3)
-      symbolic = jf1(a.reshape(vi, 3)).reshape(i).numpy()
-      expected = a.var(1).numpy()
+      # axis = 1
+      symbolic = jf1(c[:vi]).reshape(i).numpy()
+      expected = c[:i].var(1).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var_2d(self):
@@ -303,24 +294,24 @@ class TestSymbolicJit(unittest.TestCase):
     jf = TinyJit(f)
     jf0 = TinyJit(f0)
     jf1 = TinyJit(f1)
+    a = Tensor.rand(10, 10)
+    b = Tensor.rand(10, 10)
+    c = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        # aixs = None
-        a = Tensor.rand(i, j)
-        symbolic = jf(a.reshape(vi, vj)).numpy()
-        expected = a.var().numpy()
+        # axis = None
+        symbolic = jf(a[:vi, :vj]).numpy()
+        expected = a[:i, :j].var().numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 0
-        a = Tensor.rand(i, j)
-        symbolic = jf0(a.reshape(vi, vj)).reshape(j).numpy()
-        expected = a.var(0).numpy()
+        # axis = 0
+        symbolic = jf0(b[:vi, :vj]).reshape(j).numpy()
+        expected = b[:i, :j].var(0).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
-        # aixs = 1
-        a = Tensor.rand(i, j)
-        symbolic = jf1(a.reshape(vi, vj)).reshape(i).numpy()
-        expected = a.var(1).numpy()
+        # axis = 1
+        symbolic = jf1(c[:vi, :vj]).reshape(i).numpy()
+        expected = c[:i, :j].var(1).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
 if __name__ == '__main__':

test/test_symbolic_ops.py

@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, Variable
 from tinygrad.shape.shapetracker import View
-from tinygrad.helpers import Context, GlobalCounters
+from tinygrad.helpers import GlobalCounters
 from tinygrad.uop.ops import sym_infer
 from tinygrad.dtype import dtypes
 from tinygrad.device import Device
@@ -9,54 +9,46 @@ from examples.gpt2 import Attention
 import numpy as np
 
 class TestSymbolicOps(unittest.TestCase):
-  def setUp(self):
-    # A lot of these test are out of bounds, so we ignore the bounds check
-    self.context = Context(IGNORE_OOB=1)
-    self.context.__enter__()
-
-  def tearDown(self):
-    self.context.__exit__(None, None, None)
-
   def test_plus1(self):
     def f(a): return (a+1).realize()
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      symbolic = f(a.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a).numpy()
+      symbolic = f(a[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_add(self):
     def f(a, b): return (a+b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(3, i)
-      symbolic = f(a.reshape(3, vi), b.reshape(3, vi)).reshape(3, i).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b[:, :vi]).reshape(3, i).numpy()
+      expected = f(a[:, :i], b[:, :i]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_matmul(self):
     def f(a, b): return (a@b).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(10, 5)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
-      b = Tensor.rand(i, 5)
-      symbolic = f(a.reshape(3, vi), b.reshape(vi, 5)).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b[:vi, :]).numpy()
+      expected = f(a[:, :i], b[:i, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_attention(self, dropout_p=0.0, imin=1, imax=5, use_symbolic=True):
     def f(q, k, v): return Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), dropout_p=dropout_p).realize()
+    q = Tensor.rand(2, 1, 4, 8)
+    k = Tensor.rand(2, 10, 4, 8)
+    v = Tensor.rand(2, 10, 4, 8)
     for i in range(imin, imax):
       vi = Variable("i", 1, 10).bind(i) if use_symbolic else i
-      q = Tensor.rand(2, 1, 4, 8)
-      k = Tensor.rand(2, i, 4, 8)
-      v = Tensor.rand(2, i, 4, 8)
       Tensor.realize(q, k, v)
       GlobalCounters.reset()
-      symbolic = f(q, k.reshape(2, vi, 4, 8), v.reshape(2, vi, 4, 8)).reshape(2, 4, 1, 8).numpy()
-      expected = f(q, k, v).numpy()
+      symbolic = f(q, k[:, :vi, :, :], v[:, :vi, :, :]).reshape(2, 4, 1, 8).numpy()
+      expected = f(q, k[:, :i, :, :], v[:, :i, :, :]).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_attention_cmp_symbolic(self):
@@ -90,73 +82,89 @@ class TestSymbolicOps(unittest.TestCase):
 
   def test_cat_dim0(self):
     def f(a, b): return a.cat(b, dim=0).realize()
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
       b = Tensor.rand(2, 3)
-      symbolic = f(a.reshape(vi, 3), b).reshape(i+2, 3).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:vi, :], b).reshape(i+2, 3).numpy()
+      expected = f(a[:i, :], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim1(self):
     def f(a, b): return a.cat(b, dim=1).realize()
+    a = Tensor.rand(3, 10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(3, i)
       b = Tensor.rand(3, 2)
-      symbolic = f(a.reshape(3, vi), b).reshape(3, i+2).numpy()
-      expected = f(a, b).numpy()
+      symbolic = f(a[:, :vi], b).reshape(3, i+2).numpy()
+      expected = f(a[:, :i], b).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim0_two_vars(self):
     def f(a, b): return a.cat(b, dim=0).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(10, 3)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(j, 3)
-        symbolic = f(a.reshape(vi, 3), b.reshape(vj, 3)).reshape(i+j, 3).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vi, :], b[:vj, :]).reshape(i+j, 3).numpy()
+        expected = f(a[:i, :], b[:j, :]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_cat_dim1_two_vars(self):
     def f(a, b): return a.cat(b, dim=1).realize()
+    a = Tensor.rand(3, 10)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(3, i)
-        b = Tensor.rand(3, j)
-        symbolic = f(a.reshape(3, vi), b.reshape(3, vj)).reshape(3, i+j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:, :vi], b[:, :vj]).reshape(3, i+j).numpy()
+        expected = f(a[:, :i], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_two_vars_plus1_ij(self):
     def f(a, b): return (a@b+1).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(i, 3)
-        b = Tensor.rand(3, j)
-        symbolic = f(a.reshape(vi, 3), b.reshape(3, vj)).reshape(i, j).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vi, :], b[:, :vj]).reshape(i, j).numpy()
+        expected = f(a[:i, :], b[:, :j]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_two_vars_plus1_ji(self):
     # reverse the order of variables
     def f(a, b): return (a@b+1).realize()
+    a = Tensor.rand(10, 3)
+    b = Tensor.rand(3, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
-        a = Tensor.rand(j, 3)
-        b = Tensor.rand(3, i)
-        symbolic = f(a.reshape(vj, 3), b.reshape(3, vi)).reshape(j, i).numpy()
-        expected = f(a, b).numpy()
+        symbolic = f(a[:vj, :], b[:, :vi]).reshape(j, i).numpy()
+        expected = f(a[:j, :], b[:, :i]).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
+  def test_reshape_from_symbolic(self):
+    a = Tensor.rand(30)
+    for i in range(3, 5):
+      vi = Variable("i", 3, 10).bind(i)
+      symbolic = a[:vi*3].reshape((3, 3)).numpy()
+      # To match symbolic reshape (potential implicit shrink), we need a shrink
+      expected = a[:i*3].shrink(((0, 9),)).reshape((3, 3)).numpy()
+      np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
+
+  def test_invalid_symbolic_reshape(self):
+    a = Tensor.rand(30)
+    for i in range(1, 5):
+      vi = Variable("i", 1, 10).bind(i)
+      # Cannot reshape into symbolic from non-symbolic
+      with self.assertRaises(AssertionError): a.reshape((3, vi))
+
   def test_shrink(self):
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
@@ -176,11 +184,10 @@ class TestSymbolicOps(unittest.TestCase):
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_slice_no_start(self):
+    a = Tensor.rand(7, 11)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(7, 11)
-      symbolic = a[3:5, :vi:1].reshape(2,i)
-      symbolic = symbolic.numpy()
+      symbolic = a[3:5, :vi:1].reshape(2, i).numpy()
       expected = a[3:5, :i:1].numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
@@ -201,31 +208,31 @@ class TestSymbolicOps(unittest.TestCase):
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_ones_sum(self):
+    t = Tensor.ones(10)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      t = Tensor.ones(i)
-      symbolic = t.reshape(vi).sum().item()
-      expected = t.sum().item()
+      symbolic = t[:vi].sum().item()
+      expected = t[:i].sum().item()
       np.testing.assert_equal(symbolic, expected)
 
   def test_mean(self):
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
       for axis in [None, 0, 1]:
-        a = Tensor.rand(i, 3)
-        expected = a.mean(axis).numpy()
-        symbolic = a.reshape(vi, 3).mean(axis).reshape(expected.shape).numpy()
+        expected = a[:i].mean(axis).numpy()
+        symbolic = a[:vi].mean(axis).reshape(expected.shape).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_mean_2d(self):
+    a = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
         for axis in [None, 0, 1]:
-          a = Tensor.rand(i, j)
-          expected = a.mean(axis).numpy()
-          symbolic = a.reshape(vi, vj).mean(axis).reshape(expected.shape).numpy()
+          expected = a[:i, :j].mean(axis).numpy()
+          symbolic = a[:vi, :vj].mean(axis).reshape(expected.shape).numpy()
           np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var(self):
@@ -233,43 +240,43 @@ class TestSymbolicOps(unittest.TestCase):
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
       for axis in [None, 0, 1]:
-        expected = a[:i, :].var(axis).numpy()
-        symbolic = a[:vi, :].var(axis).reshape(expected.shape).numpy()
+        expected = a[:i].var(axis).numpy()
+        symbolic = a[:vi].var(axis).reshape(expected.shape).numpy()
         np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_var_2d(self):
+    a = Tensor.rand(10, 10)
     for i in range(1, 5):
       for j in range(1, 5):
         vi = Variable("i", 1, 10).bind(i)
         vj = Variable("j", 1, 10).bind(j)
         for axis in [None, 0, 1]:
-          a = Tensor.rand(i, j)
-          expected = a.var(axis).numpy()
-          symbolic = a.reshape(vi, vj).var(axis).reshape(expected.shape).numpy()
+          expected = a[:i, :j].var(axis).numpy()
+          symbolic = a[:vi, :vj].var(axis).reshape(expected.shape).numpy()
           np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=1e-6)
 
   def test_bitcast_down(self):
+    a = Tensor.rand(10, 3)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 3)
-      expected = a.bitcast(dtypes.uint8).numpy()
-      symbolic = a.reshape(vi, 3).bitcast(dtypes.uint8).reshape(expected.shape).numpy()
+      expected = a[:i].bitcast(dtypes.uint8).numpy()
+      symbolic = a[:vi].bitcast(dtypes.uint8).reshape(expected.shape).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=0)
 
   @unittest.skipIf(Device.DEFAULT == "WEBGPU", "no uint64")
   def test_bitcast_up(self):
+    a = Tensor.rand(10, 4)
     for i in range(1, 5):
       vi = Variable("i", 1, 10).bind(i)
-      a = Tensor.rand(i, 4)
-      expected = a.bitcast(dtypes.uint64).numpy()
-      symbolic = a.reshape(vi, 4).bitcast(dtypes.uint64).reshape(expected.shape).numpy()
+      expected = a[:i].bitcast(dtypes.uint64).numpy()
+      symbolic = a[:vi].bitcast(dtypes.uint64).reshape(expected.shape).numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=0)
 
   @unittest.expectedFailure
   def test_conv2d_ceildiv_edge_case(self):
     v = Variable('v', 11, 50_000)
     val = 39601
-    x = Tensor.randn(1, 22, 39601).reshape(1, 22, v.bind(val))
+    x = Tensor.randn(1, 22, 50_000)[:, :, :v.bind(val)]
     weight = Tensor.randn(256, 22, 12)
 
     result = x.conv2d(weight=weight, groups=1, stride=6, dilation=1, padding=(3, 3))

test/test_tensor.py

@@ -415,6 +415,21 @@ class TestTinygrad(unittest.TestCase):
         data = _generate_data(depth)
         np.testing.assert_allclose(Tensor(data).numpy(), np.array(data))
 
+  def test_tensor_list_implicit_cast(self):
+    data = [True, False]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-1, 0, 1, 2, 3]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+    data = [-3.5, -2.5, -1.5, 0, 1.5, 2.5, 3.5]
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.int).numpy(), torch.tensor(data, dtype=torch.int).numpy())
+    # NOTE: torch and jax raise OverflowError: Python integer -3 out of bounds for uint8
+    # np.testing.assert_equal(Tensor(data, dtype=dtypes.uint8).numpy(), torch.tensor(data, dtype=torch.uint8).numpy())
+    np.testing.assert_equal(Tensor(data, dtype=dtypes.float).numpy(), torch.tensor(data, dtype=torch.float).numpy())
+
   def test_tensor_list_special_values(self):
     if is_dtype_supported(dtypes.float16):
       data = [math.nan, -math.inf, 65504, 65519, 65519.999, 65520, 65520.1]

test/test_tensor_variable.py

@@ -1,7 +1,6 @@
 import unittest
 import numpy as np
 from tinygrad import Tensor, Variable
-from tinygrad.helpers import Context
 
 class TestTensorVariable(unittest.TestCase):
   def test_add_tvar(self):
@@ -23,43 +22,38 @@ class TestTensorVariable(unittest.TestCase):
     assert (Tensor(3) * (vv * 4)).item() == 24
 
   def test_symbolic_mean(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(2, vv)
-      ret = t.mean().item()
-      assert ret == 1
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor.ones(2, 10).contiguous()[:, :vv]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_mean_2d(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      vv2 = Variable("b", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2, vv)
-      ret = t.mean().item()
-      assert ret == 1
+    vv = Variable("a", 1, 10).bind(2)
+    vv2 = Variable("b", 1, 10).bind(2)
+    t = Tensor.ones(10, 10).contiguous()[:vv2, :vv]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_mean_2d_axis_1(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      vv2 = Variable("b", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2, vv)
-      ret = t.mean(axis=1).reshape(2, 1).numpy()
-      assert np.all(ret == 1)
+    vv = Variable("a", 1, 10).bind(2)
+    vv2 = Variable("b", 1, 10).bind(2)
+    t = Tensor.ones(10, 10).contiguous()[:vv2, :vv]
+    ret = t.mean(axis=1).reshape(2, 1).numpy()
+    assert np.all(ret == 1)
 
   def test_symbolic_mean_2d_add(self):
-    with Context(IGNORE_OOB=1):
-      add_term = Variable("c", 0, 10).bind(1)
-      vv = Variable("a", 1, 10).bind(1)
-      vv2 = Variable("b", 1, 10).bind(1)
-      t = Tensor.ones(2, 2).contiguous().reshape(vv2+add_term, vv+add_term)
-      ret = t.mean().item()
-      assert ret == 1
+    add_term = Variable("c", 0, 10).bind(1)
+    vv = Variable("a", 1, 10).bind(1)
+    vv2 = Variable("b", 1, 10).bind(1)
+    t = Tensor.ones(20, 20).contiguous()[:vv2+add_term, :vv+add_term]
+    ret = t.mean().item()
+    assert ret == 1
 
   def test_symbolic_var(self):
-    with Context(IGNORE_OOB=1):
-      vv = Variable("a", 1, 10).bind(2)
-      t = Tensor.ones(2, 2).contiguous().reshape(2, vv)
-      ret = t.var().item()
-      assert ret == 0
+    vv = Variable("a", 1, 10).bind(2)
+    t = Tensor.ones(2, 10).contiguous()[:, :vv]
+    ret = t.var().item()
+    assert ret == 0
 
   def test_symbolic_pad(self):
     vv = Variable("a", 1, 10).bind(2)
@@ -92,5 +86,15 @@ class TestTensorVariable(unittest.TestCase):
     ret = Tensor.arange(begin.bind(4), end.bind(7))
     self.assertListEqual(ret.reshape(3).tolist(), [4,5,6])
 
+  def test_variable_empty(self):
+    v = Variable("i", 1, 10)
+    # TODO: Tensor creation from unbound variable should assert
+    # with self.assertRaises(AssertionError): t = Tensor.empty(3, v)
+    vb = v.bind(3)
+    t = Tensor.empty(3, vb)
+    assert t.uop.base.buffer.size == 30
+    assert t.uop.st.shape == (3, vb)
+
+
 if __name__ == '__main__':
   unittest.main()

test/test_tiny.py

@@ -30,7 +30,10 @@ class TestTiny(unittest.TestCase):
   def test_gemm(self, N=64, out_dtype=dtypes.float):
     a = Tensor.ones(N,N).contiguous()
     b = Tensor.eye(N).contiguous()
-    self.assertListEqual((out:=a@b).flatten().tolist(), [1.0]*(N*N))
+    lst = (out:=a@b).tolist()
+    for y in range(N):
+      for x in range(N):
+        self.assertEqual(lst[y][x], 1.0, msg=f"mismatch at ({y},{x})")
     if IMAGE < 2: self.assertEqual(out.dtype, out_dtype)
 
   # *** randomness ***

test/test_uop_graph.py

@@ -445,10 +445,10 @@ class TestUOpGraph(unittest.TestCase):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), src=(), arg=0)
       v = Variable("v", 0, 20)
-      st0 = UOp(Ops.STORE, dtypes.void, src=(glbl0.index(v), UOp.const(dtypes.int, 0), UOp(Ops.IF, src=(v<16,))))
+      st0 = UOp(Ops.STORE, dtypes.void, src=(glbl0.index(v, v<16), UOp.const(dtypes.int, 0)))
       to_uops_list([st0])
 
-      st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), v, v<20))
+      st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v, v<20), v))
       with self.assertRaises(RuntimeError): to_uops_list([st1])
 
   def test_in_bounds_access_gated_local(self):
@@ -463,7 +463,7 @@ class TestUOpGraph(unittest.TestCase):
 
       gate = (gidx<400) & (lidx<8)
 
-      local_store = UOp(Ops.STORE, dtypes.void, (sbuf.index(lidx), UOp.const(dtypes.uint, 1), UOp(Ops.IF, src=(lidx<8,))))
+      local_store = UOp(Ops.STORE, dtypes.void, (sbuf.index(lidx, lidx<8), UOp.const(dtypes.uint, 1)))
 
       barrier = UOp(Ops.BARRIER, dtypes.void, (local_store,))
       if_barrier = UOp(Ops.IF, dtypes.void, (gate, barrier))
@@ -477,7 +477,7 @@ class TestUOpGraph(unittest.TestCase):
 
   def test_load_with_float_in_index(self):
     with Context(IGNORE_OOB=0):
-      ridx = UOp.range(dtypes.int, 20, 0)
+      ridx = UOp.range(20, 0)
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       i = (ridx.cast(dtypes.float)*0.68).trunc().cast(dtypes.int)
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(i, ((0<=i)&(i<16))),))
@@ -490,7 +490,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_load_cast_to_bool(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), (), 0)
-      ridx = UOp.range(dtypes.int, 20, 0)
+      ridx = UOp.range(20, 0)
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(ridx, ridx.cast(dtypes.bool).logical_not()),))
       to_uops_list([ld0])
 
@@ -499,7 +499,7 @@ class TestUOpGraph(unittest.TestCase):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       mask = UOp(Ops.DEFINE_GLOBAL, dtypes.bool.ptr(16), (), 0)
-      ridx = UOp.range(dtypes.int, 20, 0)
+      ridx = UOp.range(20, 0)
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask),)))
       to_uops_list([ld0])
 
@@ -592,8 +592,8 @@ class TestUOpGraph(unittest.TestCase):
   def test_switched_range_order(self):
     glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
     cf = UOp.const(dtypes.float, 0.0)
-    r1 = UOp.range(dtypes.int, 2, 0)
-    r2 = UOp.range(dtypes.int, 2, 1)
+    r1 = UOp.range(2, 0)
+    r2 = UOp.range(2, 1)
     alu = UOp(Ops.MUL, dtypes.int, (r2, r1))
     store = UOp(Ops.STORE, dtypes.void, (glbl.index(alu), cf))
     uops = to_uops_list([store])

test/test_uops.py

@@ -22,7 +22,7 @@ def _uops_to_prg(uops_list):
   uops = full_rewrite(ast:=UOp.sink(*uops_list), opts=Device[Device.DEFAULT].renderer)
   src = Device[Device.DEFAULT].renderer.render(uops)
   has_local = Device[Device.DEFAULT].renderer.has_local
-  return CompiledRunner(ProgramSpec("test", src, Device.DEFAULT, ast, uops=uops,
+  return CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test", src, Device.DEFAULT, ast, uops=uops,
                                 global_size=[1,1,1] if has_local else None, local_size=[1,1,1] if has_local else None))
 
 def uop(uops:list[UOp], uop:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
@@ -402,6 +402,14 @@ class TestAssembly(unittest.TestCase):
     self.assertIn(Ops.SHR, ops)
     self.assertNotIn(Ops.IDIV, ops)
 
+  def test_fast_idiv_remove_powers_of_two(self):
+    ridx = UOp.range(2**20, 0)
+    uops = to_uops_list([ridx//(7*64)], opts=Device[Device.DEFAULT].renderer)
+    ops = [x.op for x in uops]
+    # this requires shifting out the powers of two before doing fast_idiv
+    # (((ridx0>>6)*18725)>>17) instead of (int)((((long)(ridx0)*1198373)>>29))
+    self.assertNotIn(Ops.CAST, ops)
+
   def test_mulacc_unrolled(self):
     # test that     acc = acc + a0*b0 + a1*b1 + a2*b2 + a3*b3
     # is not        acc = acc + (a0*b0 + a1*b1 + a2*b2 + a3*b3)
@@ -439,7 +447,7 @@ class TestUOpMethod(unittest.TestCase):
   def test_uop_variables(self):
     a = UOp.variable("a", 1, 10)
     uop_var = Tensor(a.bind(1))
-    st_var = Tensor.empty((2, 1)).reshape((2, a.bind(1)))
+    st_var = Tensor.empty((2, 10))[:, :a.bind(1)]
     _, var_vals = (uop_var+st_var).schedule_with_vars()
     self.assertEqual(len(var_vals), 1)
     self.assertEqual(list(var_vals)[0], a)

test/unit/test_attention.py

@@ -1,5 +1,6 @@
 import unittest
-from tinygrad import Tensor, dtypes
+from tinygrad import Tensor, dtypes, TinyJit, UOp
+from tinygrad.apps.llm import apply_rope
 
 # TODO: test_scheduler, but just in uint
 class TestAttention(unittest.TestCase):
@@ -16,5 +17,29 @@ class TestAttention(unittest.TestCase):
     for si in softmax_inputs:
       assert all(b.dtype == dtypes.half for b in si.bufs), f"non half {si.bufs=}"
 
+  def test_apply_rope(self):
+    x = Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32)
+    result = apply_rope(x, 0)
+    self.assertEqual(result.shape, x.shape)
+    self.assertEqual(result.dtype, x.dtype)
+    self.assertGreater((result - apply_rope(x, 5)).abs().max().item(), 1e-6)
+    with self.assertRaises(AssertionError): apply_rope(Tensor.randn(1, 1, 4, 7, dtype=dtypes.float32), 0)
+
+  def test_apply_rope_jit_prune(self):
+    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
+    rope_noprune = TinyJit(rope_fn)
+    rope_prune = TinyJit(rope_fn, prune=True)
+
+    v_pos = UOp.variable("start_pos", 0, 100)
+    for _ in range(3):
+      rope_noprune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+      rope_prune(Tensor.randn(1, 2, 4, 8, dtype=dtypes.float32), v_pos.bind(1))
+    noprune_size = len(rope_noprune.captured.jit_cache)
+    prune_size = len(rope_prune.captured.jit_cache)
+
+    self.assertGreater(noprune_size, prune_size)
+    self.assertGreaterEqual(noprune_size, 3)
+    self.assertEqual(prune_size, 1)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_dtype.py

@@ -56,6 +56,7 @@ class TestCastConvenienceMethod(unittest.TestCase):
 class TestDtypeTolist(unittest.TestCase):
   def test_bfloat16(self):
     self.assertEqual(Tensor([-60000, 1.5, 3.1, 60000], device="PYTHON", dtype=dtypes.bfloat16).tolist(), [-59904.0, 1.5, 3.09375, 59904.0])
+  def test_fp8(self):
     # 448
     self.assertEqual(Tensor([-30000, 1.5, 3.1, 30000], device="PYTHON", dtype=dtypes.fp8e4m3).tolist(), [-448.0, 1.5, 3.0, 448.0])
     # 57344

test/unit/test_dtype_spec.py

@@ -1,6 +1,6 @@
-import unittest, math, operator, subprocess
+import unittest, math, operator, subprocess, struct
 from tinygrad.tensor import Tensor, dtypes, Device
-from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, truncate_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
+from tinygrad.dtype import DType, DTYPES_DICT, truncate, truncate_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, CI, DEBUG
 from hypothesis import given, settings, strategies as strat
@@ -26,6 +26,9 @@ def _assert_eq(tensor:Tensor, target_dtype:DType, target, tol_target_dtype:float
   except AssertionError as e:
     raise AssertionError(f"\ntensor {tensor.numpy()} dtype {tensor.dtype} does not match target {target} with dtype {target_dtype}") from e
 
+def u32_to_f32(u): return struct.unpack('f', struct.pack('I', u))[0]
+def f32_to_u32(f): return struct.unpack('I', struct.pack('f', f))[0]
+
 class TestHelpers(unittest.TestCase):
   signed_ints = (dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64)
   uints = (dtypes.uint8, dtypes.uint16, dtypes.uint32, dtypes.uint64)
@@ -102,18 +105,79 @@ class TestHelpers(unittest.TestCase):
     self.assertEqual(truncate_fp16(65504), 65504)
     self.assertEqual(truncate_fp16(65519.999), 65504)
     self.assertEqual(truncate_fp16(65520), math.inf)
+    self.assertEqual(truncate_fp16(1e-8), 0.0)
+    self.assertEqual(truncate_fp16(-65504), -65504)
+    self.assertEqual(truncate_fp16(-65519.999), -65504)
+    self.assertEqual(truncate_fp16(-65520), -math.inf)
+    self.assertTrue(math.isnan(truncate_fp16(math.nan)))
 
-  def test_truncate_bf16(self):
-    self.assertEqual(truncate_bf16(1), 1)
-    self.assertAlmostEqual(truncate_bf16(1.1), 1.09375, places=7)
-    for a in [1234, 23456, -777.777]:
-      self.assertEqual(truncate_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
-    # TODO: torch bfloat 1.1 gives 1.1015625 instead of 1.09375
+  def test_float_to_bf16(self):
+    # TODO: fuzz this better
     max_bf16 = torch.finfo(torch.bfloat16).max
-    self.assertEqual(truncate_bf16(max_bf16), max_bf16)
-    self.assertEqual(truncate_bf16(min_bf16:=-max_bf16), min_bf16)
-    self.assertEqual(truncate_bf16(max_bf16 * 1.00001), math.inf)
-    self.assertEqual(truncate_bf16(min_bf16 * 1.00001), -math.inf)
+    for a in [1, 1.1, 1234, 23456, -777.777, max_bf16, max_bf16 * 1.00001, -max_bf16, -max_bf16 * 1.00001, math.inf, -math.inf]:
+      self.assertEqual(float_to_bf16(a), torch.tensor([a], dtype=torch.bfloat16).item())
+    self.assertTrue(math.isnan(float_to_bf16(math.nan)))
+
+  def test_float_to_bf16_nan(self):
+    # In f32, NaN = exp 0xFF and mantissa ≠ 0. Quiet-vs-signaling is bit 22 of the mantissa: 1 = qNaN, 0 = sNaN.
+    # qNaN(+/-), sNaN(+/-) overflow(+/-)
+    patterns = [0x7FC00001, 0xFFC00001, 0x7F800001, 0xFF800001, 0x7FFFFFFF, 0xFFFFFFFF]
+    for u in patterns:
+      x = u32_to_f32(u)
+      y = float_to_bf16(x)
+      t = torch.tensor([x], dtype=torch.bfloat16).item()
+      self.assertTrue(math.isnan(y))
+      self.assertTrue(math.isnan(t))
+
+  def test_float_to_bf16_round(self):
+    # round_to_nearest_even
+    uppers = [0x3f800000, 0x41230000, 0xC1460000] # 1.0, 10.1875, -12.375
+    for upper in uppers:
+      base = upper & 0xFFFF0000
+      base_f32 = u32_to_f32(base)
+      base_f32_round_up = u32_to_f32(base + 0x00010000)
+
+      # low < 0x8000(0.5ULP) -> round down
+      x = u32_to_f32(base | 0x00007000)
+      self.assertEqual(float_to_bf16(x), base_f32)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+
+      # low > 0x8000(0.5ULP) -> round up
+      x = u32_to_f32(base | 0x0000C000)
+      self.assertEqual(float_to_bf16(x), base_f32_round_up)
+      self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+      # low == 0x8000(0.5ULP) and LSB even -> round down
+      if ((upper >> 16) & 1) == 0:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32)
+      # low == 0x8000(0.5ULP) and LSB odd -> round up
+      else:
+        x = u32_to_f32(base | 0x00008000)
+        self.assertEqual(float_to_bf16(x), base_f32_round_up)
+        self.assertEqual(torch.tensor([x], dtype=torch.bfloat16).item(), base_f32_round_up)
+
+  def test_float_to_bf16_boundary(self):
+    # bf16 max finite: exp=0xFE, faction=0x7F => 0x7F7F0000(f32)
+    # bf16 inf(+/-):   exp=0xFF
+    base = 0x7F7F0000
+    inf_u32 = 0x7F800000
+
+    # low < 0.5ULP
+    x = u32_to_f32(base | 0x00007FFF)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), base)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), base)
+
+    # low > 0.5ULP -> overflows to +inf
+    x = u32_to_f32(base | 0x0000C000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
+
+    # low == 0.5ULP and LSB odd -> overflows to +inf
+    x = u32_to_f32(base | 0x00008000)
+    self.assertEqual(f32_to_u32(float_to_bf16(x)), inf_u32)
+    self.assertEqual(f32_to_u32(torch.tensor([x], dtype=torch.bfloat16).item()), inf_u32)
 
   @given(strat.floats(width=32, allow_subnormal=True, allow_nan=True, allow_infinity=True))
   def test_truncate_fp8e4m3(self, x):

test/unit/test_gguf.py

@@ -53,11 +53,37 @@ class TestGGUF(unittest.TestCase):
   def test_load_tinyllama_q4_0(self): self._test_gguf_load("https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories15M-q4_0.gguf?download=true")
   def test_load_gpt2_q4_1(self): self._test_gguf_load("https://huggingface.co/PrunaAI/gpt2-GGUF-smashed/resolve/main/gpt2.Q4_1.gguf?download=true")
   def test_load_sample_q6_k(self): self._test_gguf_load("https://huggingface.co/Isotr0py/test-gguf-sample/resolve/main/Quant_Q6_K_1024.gguf?download=true")
+  def test_load_sample_mxfp4(self): self._test_gguf_load("https://huggingface.co/ngxson/boring-testing-tiny/resolve/main/stories260K-mxfp4.gguf?download=true")
 
   def test_dequantization_q4_0(self): self._test_dequantization(ggml.GGML_TYPE_Q4_0)
   def test_dequantization_q4_1(self): self._test_dequantization(ggml.GGML_TYPE_Q4_1)
   def test_dequantization_q8_0(self): self._test_dequantization(ggml.GGML_TYPE_Q8_0)
   def test_dequantization_q6_k(self): self._test_dequantization(ggml.GGML_TYPE_Q6_K)
+  def test_dequantization_mxfp4(self):
+    MXFP4 = 39
+
+    def encode(nibbles, E):
+      packed = [(low & 0xF) | ((high & 0xF) << 4) for low, high in zip(nibbles[:16], nibbles[16:])]
+      return np.array([E] + packed, dtype=np.uint8)
+
+    def decode(code, E):
+      sign = -1.0 if code * 0b1000 else 1.0
+      exp = (code >> 1) & 0b11
+      mant = code & 0b1
+      val = (1.0 + 0.5 * mant) * np.exp2(exp - 1) if exp else 0.5 * mant
+      scale = np.exp2(E - 128) if E >= 2 else np.exp2(-127 if E == 1 else -128)
+      return sign * val * scale
+
+    blocks, expected = [], []
+    rng = np.random.default_rng(42)
+    for _ in range(4):
+      E = rng.integers(0, 256)
+      codes = rng.integers(0, 16, size=32, dtype=np.uint8)
+      blocks.append(encode(codes, E))
+      expected.extend(decode(c, E) for c in codes)
+    tensor = Tensor(np.concatenate(blocks))
+    out = ggml_data_to_tensor(tensor, len(expected), MXFP4)
+    self.assertListEqual(out.numpy().tolist(), np.array(expected, dtype=np.float32).tolist())
 
   def test_expected_failure_unknown_type(self):
     with self.assertRaises(ValueError):

test/unit/test_graph_rewrite.py

@@ -65,21 +65,21 @@ class TestFoldingAndReduction(unittest.TestCase):
   def test_full_graph_rewrite_reduction_with_unused_range(self):
     const1 = UOp.const(dtypes.int32, 15)
     const2 = UOp.const(dtypes.int32, 25)
-    rng = UOp.range(dtypes.int32, 10, idx=0)
+    rng = UOp.range(10, idx=0)
     optimized_sink = apply_rewrite((const1 + const2).reduce(Ops.ADD, rng))
     expected_sum = 10 * (15 + 25)
     self.assertEqual(optimized_sink.arg, expected_sum)
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_range_reduction(self):
-    simple_range = UOp.range(dtypes.int32, 5, idx=0)
+    simple_range = UOp.range(5, idx=0)
     optimized_sink = apply_rewrite(simple_range.reduce(Ops.ADD, simple_range))
     expected_sum = sum(range(5))
     self.assertEqual(optimized_sink.arg, expected_sum)
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_simple_reduction_folding(self):
-    simple_range = UOp.range(dtypes.int32, 4, idx=0)
+    simple_range = UOp.range(4, idx=0)
     add_uop = simple_range + UOp.const(dtypes.int32, 1)
     optimized_sink = apply_rewrite(add_uop.reduce(Ops.ADD, simple_range))
     expected_sum = sum(i + 1 for i in range(4))
@@ -87,8 +87,8 @@ class TestFoldingAndReduction(unittest.TestCase):
 
   @unittest.skip("currently failing")
   def test_full_graph_rewrite_nested_loop_collapse(self):
-    outer_range = UOp.range(dtypes.int32, 8, 0)
-    inner_range = UOp.range(dtypes.int32, 4, 1)
+    outer_range = UOp.range(8, 0)
+    inner_range = UOp.range(4, 1)
     expr = (outer_range * 10) + inner_range
     optimized_reduce_uop = apply_rewrite(expr.reduce(Ops.ADD, outer_range, inner_range))
     self.assertEqual(optimized_reduce_uop.op, Ops.CONST)

test/unit/test_shapetracker.py

@@ -839,25 +839,22 @@ class TestRender(unittest.TestCase):
     self.assertEqual(idx.render(), "((ridx0*3)+ridx1)")
     self.assertEqual(valid.render(), "(ridx0<2)")
 
-class TestVariableReshape(unittest.TestCase):
-  def test_reshape(self):
-    st = ShapeTracker.from_shape((3,))
-    st = st.reshape((Variable("i", 1, 10),))
+class TestVariableShrink(unittest.TestCase):
+  def test_shrink(self):
+    st = ShapeTracker.from_shape((10,))
+    st = st.shrink(((0, Variable("i", 1, 10)),))
     assert len(st.views) == 1
 
-  def test_reshape_stride_0(self):
-    st = ShapeTracker.from_shape((3,), (0,))
-    st = st.reshape((Variable("i", 1, 10).bind(3),))
-    assert len(st.views) == 1, f"multiview {st}"
-
-  def test_reshape_bound(self):
-    st = ShapeTracker.from_shape((3,))
-    st = st.reshape((Variable("i", 1, 10).bind(3),))
+  def test_shrink_bound(self):
+    st = ShapeTracker.from_shape((10,))
+    st = st.shrink(((0, Variable("i", 1, 10).bind(3)),))
     assert len(st.views) == 1
 
-  def test_add(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10),))
+class TestVariableMerge(unittest.TestCase):
+  def test_add_reshape(self):
+    vi = Variable("i", 1, 10)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = st1+st2
     assert len(st.views) == 1
 
@@ -867,15 +864,17 @@ class TestVariableReshape(unittest.TestCase):
     st = st1+st2
     assert len(st.views) == 1, f"multiview {st}"
 
-  def test_add_bound(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10).bind(3),))
+  def test_add_reshape_bound(self):
+    vi = Variable("i", 1, 10).bind(3)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = st1+st2
     assert len(st.views) == 1
 
   def test_simplify(self):
-    st1 = ShapeTracker.from_shape((3,))
-    st2 = ShapeTracker.from_shape((Variable("i", 1, 10).bind(3),))
+    vi = Variable("i", 1, 10).bind(3)
+    st1 = ShapeTracker.from_shape((vi,))
+    st2 = ShapeTracker.from_shape((1, vi,))
     st = ShapeTracker((st1.views[0], st2.views[0]))
     st = st.simplify()
     assert len(st.views) == 1

test/unit/test_shapetracker_math.py

@@ -87,20 +87,6 @@ class TestShapeTrackerAdd(unittest.TestCase):
     assert not (st_equal(st1, st2))
 
 class TestShapeTrackerAddVariable(unittest.TestCase):
-  def test_self_add(self):
-    j = Variable("j", 0, 20).bind(10)
-    a = ShapeTracker.from_shape((10,10))
-    x = a.reshape((10, j))
-    out = x + x
-    assert out == x
-
-  def test_self_add_reshape(self):
-    j = Variable("j", 0, 20).bind(10)
-    a = ShapeTracker.from_shape((10,10))
-    x = a.reshape((10, j))
-    out = x.reshape((5, 2, j)) + x
-    assert out == x
-
   def test_merge_symbolic_views(self):
     var_i = Variable('i', 1, 10)
     var_j = Variable('i', 1, 10)

test/unit/test_simplify_valid_idx.py

@@ -19,7 +19,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
 
 def Special(expr, nmax): return UOp(Ops.SPECIAL, dtypes.int, (), (expr, nmax))
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
-def Range(n, nmax): return UOp.range(dtypes.int, nmax, n)
+def Range(n, nmax): return UOp.range(nmax, n)
 
 class TestHelpers(unittest.TestCase):
   def test_is_increasing(self):

test/unit/test_symbolic_shapetracker.py

@@ -48,11 +48,11 @@ class TestSymbolic(unittest.TestCase):
     i = Variable("i", 1, 5).bind(3)
     j = Variable("j", 1, 5).bind(3)
     k = Variable("k", 1, 5).bind(3)
-    t = Tensor.rand(3, 4).reshape(i, 4).cat(Tensor.rand(3, 4).reshape(j, 4), dim=0).cat(Tensor.rand(3, 4).reshape(k, 4), dim=0)
+    t = Tensor.rand(5, 4)[:i].cat(Tensor.rand(5, 4)[:j], dim=0).cat(Tensor.rand(5, 4)[:k], dim=0)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (i+j+k, 4))
     assert st.real_strides() == (4, 1)
-    t = Tensor.rand(3, 3).reshape(i, 3).cat(Tensor.rand(3, 3).reshape(i, 3), dim=0).cat(Tensor.rand(3, 3), dim=0)
+    t = Tensor.rand(5, 3)[:i].cat(Tensor.rand(5, 3)[:i], dim=0).cat(Tensor.rand(3, 3), dim=0)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (2*i+3, 3))
     assert st.real_strides() == (3, 1)
@@ -61,7 +61,7 @@ class TestSymbolic(unittest.TestCase):
     i = Variable("i", 1, 5).bind(4)
     j = Variable("j", 1, 5).bind(4)
     k = Variable("k", 1, 5).bind(4)
-    t = Tensor.rand(3, 4).reshape(3, i).cat(Tensor.rand(3, 4).reshape(3, j), dim=1).cat(Tensor.rand(3, 4).reshape(3, k), dim=1)
+    t = Tensor.rand(3, 5)[:, :i].cat(Tensor.rand(3, 5)[:, :j], dim=1).cat(Tensor.rand(3, 5)[:, :k], dim=1)
     st = t.uop.st
     self.assert_tuple_equal(st.shape, (3, i+j+k))
     self.assert_tuple_equal(st.real_strides(), (i+j+k, 1))
@@ -109,60 +109,44 @@ class TestShapeTrackerUnbind(unittest.TestCase):
     assert unbound_view == View.create(shape=(v, 4))
     assert var_val == {v: 3}
 
-  def test_reshape_unbind(self):
-    v = Variable("v", 1, 100)
-    bv = Variable("v", 1, 100).bind(3)
-    t = Tensor.rand(3, 4).reshape(bv, 4)
-    unbound_st, var_val = t.uop.st.unbind()
-    assert unbound_st == ShapeTracker((View.create(shape=(v, 4)),))
-    assert var_val == {v: 3}
-
   def test_shrink_unbind(self):
     v = Variable("v", 1, 100)
     bv = Variable("v", 1, 100).bind(2)
+    t = Tensor.rand(3, 4).shrink(((0,bv),(0,4)))
+    unbound_st, var_val = t.uop.st.unbind()
+    assert unbound_st == ShapeTracker((View.create(shape=(v, 4)),))
+    assert var_val == {v: 2}
     t = Tensor.rand(3, 4).shrink(((bv, bv+1), (0, 4)))
     unbound_st, var_val = t.uop.st.unbind()
     assert unbound_st == ShapeTracker((View.create(shape=(1, 4), offset=4*v),))
     assert var_val == {v: 2}
 
-class TestSymbolicReshapeFromContiguous(unittest.TestCase):
-  def test_reshape_into_symbols_simple(self):
+class TestSymbolicReshape(unittest.TestCase):
+  def test_reshape(self):
+    a = Tensor.rand(5, 4)
+    b = Tensor.rand(5, 6)
     for i in range(1, 6):
       vi = Variable("i", 1, 5).bind(i)
-      t = Tensor.rand(i, 4).reshape(vi, 4)
-      assert t.shape == (vi, 4)
-      t = Tensor.rand(i, 6).reshape(vi, 2, 3)
-      assert t.shape == (vi, 2, 3)
-
-  def test_reshape_symbols_reshape_ints(self):
-    for i in range(1, 6):
-      vi = Variable("i", 1, 5).bind(i)
-      t = Tensor.rand(i, 4).reshape(vi, 4)
-      assert t.shape == (vi, 4)
-      t = t.reshape(i, 4)
-      assert t.shape == (i, 4)
-
-  @unittest.skip("works now")
-  def test_reshape_into_symbols_bad_shape(self):
-    vi = Variable("i", 1, 10).bind(4)
-    # TODO: this never actually worked, it relied on lazy
-    #with self.assertRaises(ValueError):
-    #  Tensor.rand(4, 6).reshape(vi, 6).reshape(1, 77) # reshape to a different size new shape through symbolic shape
-    with self.assertRaises(AssertionError):
-      Tensor.rand(3, 4).reshape(3, (vi+1)) # reshape into non-Variable Node
+      ret = a[:vi]
+      ret = ret.reshape((vi, 4))
+      assert ret.shape == (vi, 4)
+      ret = b[:vi]
+      ret = ret.reshape((vi, 2, 3))
+      assert ret.shape == (vi, 2, 3)
 
   def test_two_symbol_reshape(self):
+    t = Tensor.rand(5, 5)
     for i in range(1, 6):
       for j in range(1, 6):
         vi = Variable("i", 1, 5).bind(i)
         vj = Variable("j", 1, 5).bind(j)
-        t = Tensor.rand(i, j).reshape(vi, vj)
-        assert t.shape == (vi, vj)
-        # NOTE: this is currently not allowed
-        # t = t.reshape(1, vi*vj)
-        # assert t.shape == (1, vi*vj)
-        t = t.reshape(vj, vi)
-        assert t.shape == (vj, vi)
+        ret = t[:vi, :vj]
+        ret = ret.reshape(vj, vi)
+        assert ret.shape == (vj, vi)
+        ret = ret.reshape(vi, vj)
+        assert ret.shape == (vi, vj)
+        ret = ret.reshape(1, vi*vj)
+        assert ret.shape == (1, vi*vj)
 
   def test_symbolic_mask(self):
     # taken from gpt2 single kvcache
@@ -175,41 +159,6 @@ class TestSymbolicReshapeFromContiguous(unittest.TestCase):
     new_shape = (2, (Variable('start_pos', 1, 128)+1), 16, 64)
     assert view.reshape(new_shape) is None
 
-class TestSymbolicReshapeFromNonContiguous(unittest.TestCase):
-  def test_reshape_from_const(self):
-    vi = Variable("i", 1, 5).bind(4)
-    t = Tensor.ones(3, 4).reshape(3, vi)
-    assert t.shape == (3, vi)
-    assert not t.uop.st.contiguous
-    assert len(t.uop.st.views) == 1
-
-  def test_reshape_not_allowed(self):
-    vi = Variable("i", 1, 5).bind(4)
-    with self.assertRaises(ValueError):
-      # different shape length  # TODO: cases where contractions matched might be fine
-      Tensor.ones(3, 4, 1).reshape(3, vi)
-    with self.assertRaises(ValueError):
-      # size matched, but dimensions do not match
-      Tensor.ones(4, 3).reshape(3, vi)
-
-  def test_reshape_from_padded(self):
-    vi = Variable("i", 1, 5).bind(4)
-    t = Tensor.ones(3, 4).contiguous().expand(2, 3, 4).pad(((1, 1), None, None)).shrink((None, None, (1, 3)))
-    st = t.uop.st
-    assert len(st.views) == 1
-    view = st.views[0]
-    assert view.shape == (4, 3, 2)
-    t = t.reshape(vi, 3, 2)
-    st2 = t.uop.st
-    assert len(st2.views) == 1
-    view2 = st2.views[0]
-    # check only shape changed. strides, offset, mask, contiguous remained the same
-    assert view2.shape == (vi, 3, 2)
-    assert view.strides == view2.strides == (0, 4, 1)
-    assert view.offset == view2.offset == 1
-    assert view.mask == view2.mask == ((1, 3), (0, 3), (0, 2))
-    assert not view.contiguous and not view2.contiguous
-
 class TestSymbolicExpand(unittest.TestCase):
   def test_expand_into_symbols(self):
     vi = Variable("i", 1, 5).bind(3)
@@ -220,11 +169,12 @@ class TestSymbolicExpand(unittest.TestCase):
     assert a.shape == (3, vi, vj)
 
   def test_plus_expands_constant(self):
+    a = Tensor.rand(3, 5)
     for i in range(1, 6):
       vi = Variable("i", 1, 5).bind(i)
-      a = Tensor.rand(3, i).reshape(3, vi)
-      a = a + 1
-      self.assertTupleEqual(a.shape, (3, vi))
+      ret = a[:, :vi]
+      ret = ret + 1
+      self.assertTupleEqual(ret.shape, (3, vi))
 
   def test_pad_then_expand_into_symbols(self):
     vi = Variable("i", 1, 10).bind(3)
@@ -234,6 +184,11 @@ class TestSymbolicExpand(unittest.TestCase):
     self.assertEqual(a.reshape(vi*25).shape, (vi*25,))
 
 class TestSymbolicShrink(unittest.TestCase):
+  def test_shrink_symbols_simple(self):
+    vi = Variable("i", 1, 5)
+    t = Tensor.rand(5, 5).shrink(((0, 5),(0,vi)))
+    assert t.shape == (5, vi)
+
   def test_shrink_symbols(self):
     vi = Variable("i", 1, 5)
     t = Tensor.rand(3, 5).shrink(((0, 2), (vi, vi+1)))
@@ -242,10 +197,10 @@ class TestSymbolicShrink(unittest.TestCase):
 class TestSymbolicPad(unittest.TestCase):
   def test_pad(self):
     v = Variable("v", 1, 100).bind(5)
-    t = Tensor.ones(5).reshape(v).pad(((4, 0),)).reshape(9)
-    assert t.shape == (9,)
-    st = t.uop.st
-    print(st)
+    t = Tensor.ones(100)[:v].pad(((4, 0),))
+    t = t.reshape(9)
+    assert t.tolist() == [0,0,0,0,1,1,1,1,1]
+
 
 if __name__ == '__main__':
   unittest.main()

test/unit/test_tensor_uop_representation.py

@@ -97,7 +97,7 @@ class TestTensorUopRepresentation(unittest.TestCase):
     is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.BUFFER),)))
     vi = UOp.variable("i", 1, 3).bind(1)
     a = Tensor.empty(3, vi)
-    is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.BUFFER),)))
+    is_pattern(a, UPat(Ops.RESHAPE, src=(UPat(Ops.SHRINK, src=(UPat(Ops.BUFFER),))),))
     self.assertEqual(a.uop.base.buffer.size, 9)
 
 if __name__ == '__main__':

test/unit/test_uop_spec.py

@@ -81,5 +81,16 @@ class TestUOpSpec(unittest.TestCase):
     with self.assertRaisesRegex(RuntimeError, "UOp verification failed"):
       type_verify([a], tensor_uop_spec)
 
+class TestUOpSink(unittest.TestCase):
+  def test_0(self):
+    s = UOp.sink()
+    self.assertEqual(len(s.src), 0)
+
+  def test_1(self):
+    a = UOp.const(dtypes.int, 0)
+    s1 = UOp.sink(a)
+    s2 = a.sink()
+    self.assertIs(s1, s2)
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_uop_symbolic.py

@@ -8,7 +8,7 @@ from tinygrad.codegen.late.devectorizer import sym
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad import Variable
-from tinygrad.uop.spec import z3_renderer
+from tinygrad.uop.spec import uops_to_z3
 
 def render(self) -> tuple[str, ConstType, ConstType]:
   # NOTE: we need STORE so the ALU op has children
@@ -32,9 +32,8 @@ class TestSymbolic(unittest.TestCase):
   def helper_test_variable(self, v, n, m, s, test_z3:bool=True):
     if test_z3:
       solver = z3.Solver()
-      z3_sink = graph_rewrite(v.sink(v.simplify()), z3_renderer, ctx=(solver, {}))
-      expr, epxr_simplified = z3_sink.src[0].arg, z3_sink.src[1].arg
-      self.assertEqual(solver.check(expr != epxr_simplified), z3.unsat, "simplified expression not equal to original")
+      expr, expr_simplified = uops_to_z3(solver, v, v.simplify())
+      self.assertEqual(solver.check(expr != expr_simplified), z3.unsat, "simplified expression not equal to original")
     rendered, nmin, nmax = render(v)
     if isinstance(s, tuple): self.assertIn(rendered, s)
     else: self.assertEqual(rendered, s)
@@ -640,15 +639,16 @@ class TestSymbolic(unittest.TestCase):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
     b = Variable("b", 0, 3)
+    c = Variable("c", 0, 3)
     aa = cond.where(a, a.ufix(0))
     bb = cond.where(b, b.ufix(1))
     self.helper_test_variable(aa, 0, 3, "(a if (x<2) else 0)")
     self.helper_test_variable(bb, 0, 3, "(b if (x<2) else 1)")
     self.helper_test_variable(aa+bb, 0, 6, "((a+b) if (x<2) else 1)")
     self.helper_test_variable(aa.maximum(bb), 0, 3, "(max(a, b) if (x<2) else 1)")
+    self.helper_test_variable((c+aa)+bb, 0, 9, "(c+((a+b) if (x<2) else 1))")
 
     # not combining because it increased total ALU
-    c = Variable("c", 0, 3)
     cc = cond.where(c, c+1)
     self.helper_test_variable(bb+cc, 0, 7, "((b if (x<2) else 1)+(c if (x<2) else (c+1)))")
 

test/unit/test_viz.py

@@ -281,10 +281,10 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
         v["shapes"].append({"name":strings[name], "ref":option(ref), "st":st, "dur":dur, "cat":option(cat)})
     else:
       v["peak"] = u("<Q")[0]
-      v["timestamps"] = list(u(f"<{u('I')[0]}I"))
       for _ in range(event_count):
-        i = u("<I")[0]
-        v["shapes"].append({"x":list(u(f"<{i}I")), "y":list(u(f"<{i}Q")), "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+        alloc, ts, key = u("<BII")
+        if alloc: v["shapes"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
+        else: v["shapes"].append({"event":"free", "ts":ts, "key":key})
   return {"dur":dur, "peak":global_peak, "layout":layout}
 
 class TestVizProfiler(unittest.TestCase):
@@ -376,8 +376,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{a.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 2])
+    self.assertEqual(len(ret["shapes"]), 2)
 
   def test_del_once(self):
     a = _alloc(1)
@@ -386,10 +385,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{b.device} Memory"]
     self.assertEqual(ret["peak"], 1)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 2])
-    self.assertEqual(ret["shapes"][1]["x"], [2, 3])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [0, 0])
+    self.assertEqual(len(ret["shapes"]), 3)
 
   def test_alloc_free(self):
     a = _alloc(1)
@@ -399,12 +395,7 @@ class TestVizMemoryLayout(BaseTestViz):
     profile_ret = load_profile(Buffer.profile_events)
     ret = profile_ret["layout"][f"{c.device} Memory"]
     self.assertEqual(ret["peak"], 2)
-    self.assertEqual(ret["shapes"][0]["x"], [0, 3])
-    self.assertEqual(ret["shapes"][1]["x"], [1, 3, 3, 4])
-    self.assertEqual(ret["shapes"][0]["y"], [0, 0])
-    self.assertEqual(ret["shapes"][1]["y"], [1, 1, 0, 0])
-    self.assertEqual(ret["shapes"][2]["x"], [3, 4])
-    self.assertEqual(ret["shapes"][2]["y"], [1, 1])
+    self.assertEqual(len(ret["shapes"]), 4)
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/apps/llm.py

@@ -53,17 +53,15 @@ class SimpleTokenizer:
     try: return [ self._normal_tokens[p] for p in parts ]
     except KeyError: raise RuntimeError("token not found")
 
-def apply_rope(x:Tensor, start_pos:int|UOp, base:int=10000):
+def apply_rope(x:Tensor, start_pos:int|UOp, base:float = 10000.0) -> Tensor:
   B, H, T, Hd = x.shape
-  # NOTE: this is usually in a RoPE cache, but tinygrad JIT should prune it outside the kernel
-  # TODO: make it do that
-  freq = base ** (-Tensor.arange(0, 1, 2/Hd, dtype='float32'))
-  angles = Tensor.arange(start_pos, start_pos+T, dtype='float32')[None, None, :, None] * freq
-  cos, sin = angles.cos(), angles.sin()
-  x = x.reshape(B, H, T, Hd // 2, 2)    # split into pairs
-  y1 = x[..., 0] * cos - x[..., 1] * sin
-  y2 = x[..., 0] * sin + x[..., 1] * cos
-  return Tensor.stack(y1, y2, dim=-1).reshape(B, H, T, Hd)
+  assert (Hd & 1) == 0, "RoPE requires an even head dimension"
+  half = Hd // 2
+  angles = (Tensor.arange(T, dtype="float32") + start_pos)[:, None] * (base ** (-(Tensor.arange(half, dtype="float32") / half)))[None, :]
+  cos, sin = angles.cos().reshape(1, 1, T, half).cast(x.dtype), angles.sin().reshape(1, 1, T, half).cast(x.dtype)
+  x_pairs = x.reshape(B, H, T, half, 2)
+  return Tensor.stack(x_pairs[..., 0] * cos - x_pairs[..., 1] * sin,
+                      x_pairs[..., 0] * sin + x_pairs[..., 1] * cos, dim=-1).reshape(B, H, T, Hd)
 
 class TransformerBlock:
   def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int, norm_eps:float, max_context:int=0):

tinygrad/codegen/__init__.py

@@ -1,7 +1,7 @@
 from typing import Any, Callable
 import functools
 from dataclasses import dataclass
-from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL
+from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL, RANGEIFY, POSTOPT
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp
 from tinygrad.uop.spec import type_verify
 from tinygrad.renderer import Renderer
@@ -12,12 +12,14 @@ from tinygrad.codegen.quantize import pm_quant
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing
 from tinygrad.uop.decompositions import get_late_rewrite_patterns
-from tinygrad.codegen.late.expander import migrate_indexing, expander
+from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
 from tinygrad.codegen.late.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
 from tinygrad.codegen.opt import pm_get_optimization, pm_do_optimize
 from tinygrad.codegen.opt.swizzler import view_left, view_right, fix_kernel_ops
+from tinygrad.codegen.opt.postrange import pm_postrange_opt
+from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen
 
 @dataclass
 class RewriteStep:
@@ -44,10 +46,10 @@ rewrites_for_linearizer = [
 
 def get_rewrites_for_renderer(opts:Renderer, linearizer:bool=True) -> list[RewriteStep]:
   # cache with the values of the context vars
-  return _get_rewrites_for_renderer(opts, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value)
+  return _get_rewrites_for_renderer(opts, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value, RANGEIFY.value, POSTOPT.value)
 
 @functools.cache
-def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL) -> list[RewriteStep]:
+def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL, _RANGEIFY, _POSTOPT) -> list[RewriteStep]:
   # ** lowerer (rewrite_shapetracker_with_index) **
   ret: list[RewriteStep] = []
 
@@ -55,25 +57,30 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   ret.extend(rewrites_for_views)
 
   # this is kernel.py
-  ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
-  ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
+  if not _RANGEIFY: ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
+  if not _POSTOPT and not _RANGEIFY: ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
 
   if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
   ret.append(RewriteStep(pm_lowerer, get_index, name="lowerer", bottom_up=True))
 
+  if _POSTOPT or _RANGEIFY: ret.append(RewriteStep(pm_postrange_opt, ctx=lambda _: opts, name="post optimize ast"))
+
   # ** expander (expand_rewrite) **
   ret.append(RewriteStep(sym+migrate_indexing, name="initial symbolic"))
 
-  # add gpu dims (late). this also handles UNROLL range
-  ret.append(RewriteStep(pm_add_gpudims, lambda _: opts, name="add gpudims"))
-
   # expand
-  ret.append(RewriteStep(sym+expander, name="expander"))
+  ret.append(RewriteStep(sym+pm_pre_expander+expander, name="expander"))
+
+  # add locals
+  ret.append(RewriteStep(pm_add_buffers_local+rangeify_codegen, name="add local buffers"))
 
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
   ret.append(RewriteStep(pm_reduce+gep_pushing, lambda _: ReduceContext(), name="remove_reduce"))
 
+  # add gpu dims (late). this works after devectorize, but it's faster here
+  ret.append(RewriteStep(pm_add_gpudims, lambda _: opts, name="add gpudims"))
+
   # devectorize (TODO: does this need opts?)
   if _DEVECTORIZE >= 2: pm_devectorize = sym+load_store_folding+load_store_indexing
   elif _DEVECTORIZE: pm_devectorize = sym+devectorize+load_store_folding+correct_load_store+load_store_indexing

tinygrad/codegen/gpudims.py

@@ -1,6 +1,6 @@
 import math
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType
-from tinygrad.helpers import all_int, partition, flatten, prod, dedup
+from tinygrad.helpers import all_int, dedup
 from tinygrad.dtype import dtypes
 from tinygrad.shape.view import get_contraction
 from tinygrad.renderer import Renderer
@@ -56,17 +56,17 @@ def add_gpudims(ctx:Renderer, s:UOp):
   if any(x.op is Ops.SPECIAL for x in s_topo): return None
 
   # get ranges
-  all_ranges = {x.arg[0]%1000:x for x in s_topo if x.op is Ops.RANGE}
+  all_ranges = {x.arg[0:-1]:x for x in s_topo if x.op is Ops.RANGE}
 
   # extract global/local dims
-  global_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[1] is AxisType.GLOBAL]))
-  local_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[1] in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]))
+  global_dims = sorted(dedup([x.arg[0:-1] for x in all_ranges.values() if x.arg[-1] is AxisType.GLOBAL]))
+  local_dims = sorted(dedup([x.arg[0:-1] for x in all_ranges.values() if x.arg[-1] in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]))
   if not global_dims and not local_dims: return None
 
   # get global and local shape
   ranges = [all_ranges[r] for r in global_dims+local_dims if r in all_ranges]
-  global_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0]%1000 in global_dims])
-  local_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0]%1000 in local_dims])
+  global_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0:-1] in global_dims])
+  local_shape = tuple([ssimplify(r.src[0]) for r in ranges if r.arg[0:-1] in local_dims])
 
   # get the idxs
   ki: KernelInfo = s.arg
@@ -82,34 +82,13 @@ def add_gpudims(ctx:Renderer, s:UOp):
   for r in s_topo:
     if r.op is not Ops.RANGE: continue
     try:
-      ii = (global_dims+local_dims).index(r.arg[0]%1000)
-      if r.arg[0] < 2000 and r.arg[1] == AxisType.GROUP_REDUCE: continue
+      ii = (global_dims+local_dims).index(r.arg[0:-1])
+      if r.arg[1] == AxisType.REDUCE: continue
       subs[r] = idxs[ii]
     except ValueError: continue
   return s.substitute(subs)
 
-def fix_reduce_unroll(x:UOp):
-  reduce_range, reduce_expand = partition(x.src[1:], lambda y: y.op is Ops.RANGE)
-  if len(reduce_expand) == 0: return None
-  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand} for {x.axis_arg}"
-  ret = x.src[0]
-  if len(contract_axis:=flatten(x.arg for x in reduce_expand)):
-    ret = UOp(Ops.CONTRACT, x.dtype.vec(prod(x[1] for x in contract_axis)), (ret,), tuple(contract_axis), tag=1)
-  # REDUCE supports both "horizontal" reduction and range reduction. the horizontal elements are taken in the nearest group
-  return x.replace(src=(ret,)+tuple(reduce_range))
-
-def fix_store_unroll(x:UOp):
-  store_expand, store_range = partition(x.src[2:], lambda y: y.op is Ops.UNROLL)
-  if len(store_expand) == 0: return None
-  return UOp(Ops.CONTRACT, dtypes.void, (x.replace(src=x.src[:2]+tuple(store_range)),), tuple(flatten(x.arg for x in store_expand)), tag=1)
-
 pm_add_gpudims = PatternMatcher([
+  # add gpudims must be last
   (UPat(Ops.SINK, name="s"), add_gpudims),
-  # rewrite UPCAST/UNROLL range to something to be expanded
-  (UPat(Ops.RANGE, name="r"),
-   lambda r: UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(s:=r.vmax+1), tuple(range(s))),), ((r.arg[0],s),)) \
-    if r.arg[1] in {AxisType.UNROLL, AxisType.UPCAST} else None),
-  # fix REDUCEs with UNROLLs
-  (UPat(Ops.REDUCE, name="x"), fix_reduce_unroll),
-  (UPat(Ops.STORE, name="x"), fix_store_unroll),
 ])

tinygrad/codegen/late/devectorizer.py

@@ -232,17 +232,21 @@ def no_vectorized_alu(alu:UOp):
   alus = tuple(UOp(alu.op, alu.dtype.scalar(), tuple(s.gep(i) for s in alu.src), alu.arg) for i in range(alu.dtype.vcount))
   return UOp(Ops.VECTORIZE, alu.dtype, alus)
 
-def no_vectorized_acc(acc:UOp, c:UOp):
-  if acc.dtype.count == 1: return None
-  assert c.arg == 0, "this only supports index 0"
-  new_acc = acc.replace(dtype=acc.dtype.base.scalar().ptr(acc.dtype.count, cast(PtrDType, acc.dtype).addrspace))
-  return UOp(Ops.PTRCAT, acc.dtype, tuple([new_acc.index(UOp.const(dtypes.int, i)) for i in range(acc.dtype.count)]))
+def no_vectorized_buf(buf:UOp):
+  dtype = cast(PtrDType, buf.dtype)
+  return buf.replace(dtype=dtype.base.scalar().ptr(dtype.size*dtype.count, dtype.addrspace)).cast(dtype)
+
+def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
+  cnt = cast.dtype.count
+  assert idx.dtype.count == 1, f"idx dtype must be 1 {idx.dtype}"
+  return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.int.vec(cnt), tuple(range(cnt))))
 
 devectorize = PatternMatcher([
   # no ALU on vectorized dtypes
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name="alu"), no_vectorized_alu),
   (UPat(Ops.WMMA, name="wmma"), no_vectorized_wmma),
-  (UPat(Ops.DEFINE_REG, name="acc").index(UPat.cvar("c")), no_vectorized_acc),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
 ])
 
 pm_render = PatternMatcher([

tinygrad/codegen/late/expander.py

@@ -1,9 +1,9 @@
 # this converts a lowerer program into a vectorized program
 
 import functools, itertools, operator
-from tinygrad.dtype import dtypes
-from tinygrad.helpers import AMX, dedup, flatten, all_same, prod
-from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp
+from tinygrad.dtype import dtypes, PtrDType, AddrSpace
+from tinygrad.helpers import AMX, dedup, flatten, all_same, prod, partition
+from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp, AxisType
 
 def _expand_arg_to_idx(args:tuple[tuple[int, int], ...], rpk:dict[int, int]) -> int:
   idx, mul = 0, 1
@@ -50,9 +50,11 @@ def do_expand(root:UOp):
       if root.op is Ops.IF or src.op is Ops.IF:
         # for the first arg of IF, just pass them through ignoring UNROLLS
         new_srcs.append(src)
-      elif (root.op is Ops.STORE and i >= 2) or (root.op is Ops.REDUCE and i >= 1):
+      elif (root.op is Ops.STORE and i >= 2) or (root.op in {Ops.REDUCE, Ops.BUFFERIZE} and i >= 1) or (root.op is Ops.WMMA and i >= 3):
         # for any range args of STORE/REDUCE, pass them through
         new_srcs.append(src)
+      elif root.op is Ops.INDEX and i >= 1 and not isinstance(root.dtype, PtrDType):
+        new_srcs.append(src)
       elif src.dtype.count > 1:
         # put any input dtype > 1 grouped together
         new_srcs.append(UOp(Ops.CAT, src.dtype.scalar().vec(expand_sz*src.dtype.count), (src,)*expand_sz))
@@ -84,7 +86,7 @@ expander = PatternMatcher([
   (UPat(Ops.UNROLL, name="outer", src=(UPat(Ops.UNROLL, name="inner"),)),
    lambda outer, inner: UOp(Ops.UNROLL, outer.dtype, (inner.src[0],), inner.arg+outer.arg)),
   # do expansion
-  (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX,
+  (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX, Ops.BUFFERIZE,
          Ops.VECTORIZE, Ops.IF, Ops.REDUCE), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded
@@ -112,3 +114,49 @@ migrate_indexing = PatternMatcher([
   # create gate MUST BE BEFORE expander
   (UPat(Ops.STORE, name="root"), create_gate),
 ])
+
+# ****
+
+def fix_reduce_unroll(x:UOp):
+  reduce_range, reduce_expand = partition(x.src[1:], lambda y: y.op is Ops.RANGE)
+  if len(reduce_expand) == 0: return None
+  reduce_expand = [x for x in reduce_expand if x.op is not Ops.CONST]
+  assert all(x.op is Ops.UNROLL for x in reduce_expand), f"not all UNROLLS in {reduce_expand}"
+  ret = x.src[0]
+  if len(contract_axis:=flatten(x.arg for x in reduce_expand)):
+    ret = UOp(Ops.CONTRACT, x.dtype.vec(prod(x[1] for x in contract_axis)), (ret,), tuple(contract_axis), tag=1)
+  # REDUCE supports both "horizontal" reduction and range reduction. the horizontal elements are taken in the nearest group
+  return x.replace(src=(ret,)+tuple(reduce_range))
+
+def fix_store_unroll(x:UOp):
+  store_expand, store_range = partition(x.src[2:], lambda y: y.op is Ops.UNROLL)
+  if len(store_expand) == 0: return None
+  return UOp(Ops.CONTRACT, dtypes.void, (x.replace(src=x.src[:2]+tuple(store_range)),), tuple(flatten(x.arg for x in store_expand)), tag=1)
+
+def fix_group_for_reduce(x:UOp):
+  reduce_gfr, reduce_r = partition(x.src[1:], lambda u: u.op is Ops.RANGE and u.arg[1] == AxisType.GROUP_REDUCE)
+  if len(reduce_gfr) == 0: return None
+
+  # NOTE: if there's other locals here, we need them in the buffer too
+  upstream_locals = [u for u in x.toposort() if u.op is Ops.RANGE and u.arg[1] == AxisType.LOCAL]
+
+  # do only the non grouped reduces early
+  ret = x.replace(src=(x.src[0],)+tuple(reduce_r))
+  reduce_loop = [x.replace(arg=(x.arg[0]+100, AxisType.REDUCE)) for x in reduce_gfr]
+  buf = ret.bufferize(*upstream_locals, *reduce_gfr, arg=(AddrSpace.LOCAL, reduce_gfr[0].arg[0])).index(*upstream_locals, *reduce_loop)
+
+  # gate with an if on the store + do the final reduce
+  buf = UOp(Ops.IF, dtype=buf.dtype, src=(functools.reduce(operator.and_, [x.eq(0) for x in reduce_gfr]), buf))
+  return buf.reduce(*reduce_loop, arg=x.arg)
+
+pm_pre_expander = PatternMatcher([
+  # rewrite UPCAST/UNROLL range to something to be expanded
+  (UPat(Ops.RANGE, name="r"),
+   lambda r: UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(s:=r.vmax+1), tuple(range(s))),), ((r.arg[0],s),)) \
+    if r.arg[1] in {AxisType.UNROLL, AxisType.UPCAST} else None),
+  # fix REDUCEs with UNROLLs
+  (UPat(Ops.REDUCE, name="x"), fix_reduce_unroll),
+  (UPat(Ops.STORE, name="x"), fix_store_unroll),
+  # fix group for reduce
+  (UPat(Ops.REDUCE, name="x"), fix_group_for_reduce),
+])

tinygrad/codegen/lowerer.py

@@ -1,9 +1,6 @@
 # the job of the lowerer is to do indexing
-import functools, operator
-from typing import cast
 from dataclasses import dataclass
-from tinygrad.dtype import dtypes, AddrSpace, PtrDType
-from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite
+from tinygrad.uop.ops import KernelInfo, UOp, Ops, PatternMatcher, UPat, sint_to_uop, AxisType, graph_rewrite, resolve
 
 # ***** indexing *****
 
@@ -14,12 +11,12 @@ class IndexContext:
   start: int = 0
 
 def shape_to_idx(s, axis_types, start=0):
-  return [UOp.range(dtypes.int, sint_to_uop(s), start+i, axistype=at) for i, (s, at) in enumerate(zip(s, axis_types))]
+  return [UOp.range(sint_to_uop(s), start+i, at) for i, (s, at) in enumerate(zip(s, axis_types))]
 
 def get_index(ast:UOp) -> IndexContext:
   axis_types = ast.arg.axis_types if isinstance(ast.arg, KernelInfo) else ()
-  if len(ast.full_shape) != len(axis_types):
-    axis_types = tuple([AxisType.REDUCE if s is not fs else AxisType.LOOP for s,fs in zip(ast.shape, ast.full_shape)])
+  if len(ast.full_shape) != len(axis_types) and ast.st is not None:
+    axis_types = tuple([AxisType.REDUCE if resolve(s != fs) else AxisType.LOOP for s,fs in zip(ast.shape, ast.full_shape)])
   return IndexContext(axis_types, [], 0)
 
 # ***** lowering (given index) *****
@@ -50,15 +47,7 @@ def lower_store(ctx: IndexContext, x: UOp, buf: UOp):
 
   stored = subblock(ctx, real_new_idxs, x.src[1])
   used_ranges = [x for x in used_idxs if x.op is Ops.RANGE]
-  ret = buf.index(idx, valid).store(stored, *used_ranges)
-
-  # insert BARRIER if we are ending a LOCAL, IF if we are ending a GROUP_REDUCE
-  if cast(PtrDType, buf.dtype).addrspace == AddrSpace.LOCAL and \
-      any(ctx.axis_types[x.arg[0]%1000] in {AxisType.GROUP_REDUCE, AxisType.LOCAL} for x in used_ranges):
-    ret = ret.barrier()
-    range_gates = [x.eq(0) for x in used_ranges if ctx.axis_types[x.arg[0]%1000] == AxisType.GROUP_REDUCE]
-    if len(range_gates): ret = UOp(Ops.IF, src=(functools.reduce(operator.and_, range_gates), ret))
-  return ret
+  return buf.index(idx, valid).store(stored, *used_ranges)
 
 def fixup_wmma(ctx:IndexContext, x:UOp):
   if x.tag is not None: return None

tinygrad/codegen/opt/__init__.py

@@ -7,7 +7,7 @@ from tinygrad.helpers import NOOPT, BEAM, USE_TC, getenv
 from tinygrad.renderer import Renderer
 from tinygrad.uop.spec import type_verify
 
-def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp:
+def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp|None:
   """
   Optimize an AST based on heuristics or BEAM search.
 
@@ -19,19 +19,18 @@ def get_optimized_ast(ast:UOp, renderer:Renderer) -> UOp:
     The Ops.SINK rooted AST transformed to apply the opts and with a KernelInfo in the arg.
   """
 
-  assert ast.arg is None, "no opt if there's an arg"
-  k = Kernel(ast, opts=renderer)
-  if not NOOPT:
+  # no shape, no opt
+  if ast.src[0].st is None: return None
+  new_arg = ast.arg
+  if new_arg is None and not NOOPT and not BEAM:
+    k = Kernel(ast, opts=renderer)
     if not k.apply_tensor_cores(USE_TC.value): k.apply_opts(hand_coded_optimizations(k))
-    if BEAM >= 1:
-      from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
-      kb = Kernel(ast, opts=renderer)
-      rawbufs = bufs_from_lin(kb, allocate=False)
-      k = beam_search(kb, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
-  return ast.replace(arg=KernelInfo(opts_to_apply=tuple(k.applied_opts)))
+    new_arg = KernelInfo(opts_to_apply=tuple(k.applied_opts))
+  elif new_arg is not None and len(new_arg.applied_opts): return None
+  return Kernel(ast.replace(arg=None), opts=renderer).get_optimized_ast().replace(arg=new_arg)
 
 pm_get_optimization = PatternMatcher([
-  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: get_optimized_ast(ast, ctx) if ast.arg is None and ast.src[0].st is not None else None),
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: get_optimized_ast(ast, ctx)),
 ])
 
 def apply_opt(ast:UOp, renderer:Renderer):

tinygrad/codegen/opt/kernel.py

@@ -10,7 +10,7 @@ from tinygrad.uop.spec import type_verify, ast_spec
 from tinygrad.device import Device
 from tinygrad.codegen.opt.tc import TensorCore
 from tinygrad.renderer import Renderer
-from tinygrad.dtype import ImageDType, AddrSpace
+from tinygrad.dtype import ImageDType
 from tinygrad.helpers import all_same, colored, ansilen, dedup, prod, round_up, to_function_name, unwrap, argfix, DEBUG, TC_SELECT, TC_OPT, AMX
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import strides_for_shape, get_contraction
@@ -60,7 +60,7 @@ class Kernel:
 
     self.vars: list[Variable] = self.ast.variables()
     # NOTE: this requires a specific order with the [::-1], this is likely a bug
-    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer][::-1]
+    self.bufs: list[UOp] = [x for x in self.ast.toposort() if x.op in GroupOp.Buffer and x.st is not None][::-1]
 
     # create new shapetrackers inside this kernel, we will permute them
     self.sts: list[ShapeTracker] = [x.st_arg for x in self.bufs]
@@ -92,10 +92,6 @@ class Kernel:
     global_loops = AxisType.GLOBAL if self.opts.has_local else AxisType.LOOP
     self.axis_types: list[AxisType] = [AxisType.REDUCE if resolve(x!=y) else global_loops for x,y in zip(self.output_shape, self.full_shape)]
 
-    # confirm all reduce axes are at the end
-    if (final_reduces := [x for x in self.axis_types if x == AxisType.REDUCE]) and final_reduces != self.axis_types[-len(final_reduces):]:
-      raise RuntimeError(f"reduces are not at the end of the shape {self.full_shape} -> {self.output_shape}")
-
   def copy(self):
     ret = type(self).__new__(type(self))
 
@@ -122,7 +118,7 @@ class Kernel:
   @property
   def output_shape(self) -> tuple[sint, ...]: return self.sts[0].shape
   @property
-  def shape_len(self) -> int: return len(self.sts[0].shape)
+  def shape_len(self) -> int: return len(self.full_shape)
 
   def axes_of(self, *axis_type:AxisType) -> list[int]: return [i for i,t in enumerate(self.axis_types) if t in argfix(axis_type)]
   @property
@@ -174,7 +170,7 @@ class Kernel:
   # amount : the amount to take
   # top : if you want to pull that amount from the top
   # insert_at : place to insert the new stuff
-  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None):
+  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None) -> int:
     if insert_at is None: insert_at = self.shape_len
     self.axis_types.insert(insert_at, new_type)
     move_axis = axis if top else axis+1
@@ -183,6 +179,7 @@ class Kernel:
     new_axes = [i for i in range(insert_at) if i != move_axis]+[move_axis]+[i for i in range(insert_at, self.shape_len+1) if i != move_axis]
     self.reshape(new_shape_fxn)
     self.permute(new_axes)
+    return insert_at
 
   # ******************** complex simplifiers ********************
 
@@ -244,11 +241,11 @@ class Kernel:
       if axis is None: return -1
       if op is OptOps.UNROLL: return self.unrollable_dims[axis]
       if op in {OptOps.GROUP, OptOps.GROUPTOP}: return self.axes_of(AxisType.REDUCE)[axis]
-      check(axis < self.shape_len, "invalid axis")
+      check(axis < self.shape_len, f"invalid axis on {axis=} {op=} {self.shape_len=}")
       return axis
     except IndexError as e: raise KernelOptError from e
 
-  def apply_opt(self, opt:Opt, append_opt:bool=True):
+  def apply_opt(self, opt:Opt, append_opt:bool=True) -> int|None:
     if self.finalized: raise RuntimeError("can't optimize Kernel after it's finalized")
     if self.dont_use_locals: check(opt.op not in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}, "not using locals")
 
@@ -262,7 +259,7 @@ class Kernel:
       check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
       check(self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt), "no tensor core available")
       self.applied_opts.append(opt)
-      return
+      return None
 
     axis = self.real_axis(opt.op, opt.axis)
 
@@ -285,28 +282,30 @@ class Kernel:
       smem_sz = amt*acc_sz*upcast_sz*local_sz
       check(smem_sz <= self.opts.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.opts.shared_max}")
 
+    new_axis = None
     if opt.op is OptOps.LOCAL:    # cyan
       # NOTE: LLVM/CPU can use locals too, but they are treated the same as globals (still helpful for L1 cache)
       # it's disabled for now since it makes BEAM slow for little gain
       check(self.opts.has_local, "target does not support local")
       check(self.axis_types[axis] is AxisType.GLOBAL, "local is for globals")
-      self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.LOCAL, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL))+1)
     elif opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:   # green
       check(self.opts.has_local and self.opts.has_shared, "target does not support local or shared mem")
       check(self.axis_types[axis] is AxisType.REDUCE, "must be reduce axis to group")
       check(not self.tensor_core, "can't group with tensor cores")
       check(len(reduce_axes:=[i for r in self.reduceops for i in r.axis_arg]) == len(set(reduce_axes)), "can't group with parallel reduces")
-      self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
+      new_axis = self.shift_to(axis, amt, AxisType.GROUP_REDUCE, top=(opt.op is OptOps.GROUPTOP), insert_at=min(self.axes_of(AxisType.REDUCE)))
     elif opt.op is OptOps.UNROLL:                     # purple
       check(self.axis_types[axis] not in (AxisType.UPCAST, AxisType.UNROLL), "can't upcasted already upcasted")
       check(amt <= 32, "don't unroll more than 32")
-      self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
+      new_axis = self.shift_to(axis, amt, AxisType.UNROLL, insert_at=None)
     elif opt.op is OptOps.UPCAST:                     # yellow
       check(axis in self.upcastable_dims, f"{axis=} not in {self.upcastable_dims=}")
       # NOTE: assume the first get_local_axes() LOCAL are for TC
       check(not (self.tensor_core and axis in self.axes_of(AxisType.LOCAL)[:len(self.tensor_core.get_local_axes())]), "can't upcast TC locals")
       check((self.opts is not None and self.opts.device == "DSP") or amt <= 16, "don't upcast more than 16")
-      self.shift_to(axis, amt, AxisType.UPCAST, insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
+      new_axis = self.shift_to(axis, amt, AxisType.UPCAST,
+                               insert_at=max(self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP, AxisType.UPCAST))+1)
     elif opt.op is OptOps.NOLOCALS:
       check(self.opts.has_local and not self.dont_use_locals, "NOLOCALS is meaningless if target does not support local or already not using locals")
       check(AxisType.LOCAL not in self.axis_types and self.group_for_reduces == 0, "can't have no locals with locals")
@@ -336,6 +335,7 @@ class Kernel:
     if append_opt: self.applied_opts.append(opt)
     if self.simplify_ones() and self.tensor_core_opts:
       self.tensor_core_opts.fix_axes(axis) # fix up axes in TC opts if required after simplify_ones()
+    return new_axis
 
   def apply_opts(self, opts:Sequence[Opt]) -> Kernel:
     for opt in opts: self.apply_opt(opt)
@@ -445,6 +445,7 @@ class Kernel:
   def shape_str_to_axis(self, nms:list[str]) -> tuple[int, ...]: return tuple([self.shape_str().index(x) for x in nms])
 
   def get_optimized_ast(self, name_override:str|None=None) -> UOp:
+    if self.applied_opts: raise RuntimeError("not supported")
     @functools.cache
     def fixup_ast(op:UOp) -> UOp:
       ret = op.replace(src=tuple(fixup_ast(x) for x in op.src)) # noqa: F821
@@ -460,8 +461,7 @@ class Kernel:
       if op.op is Ops.REDUCE_AXIS:
         reduce_idx = len(self.bufs) + self.reduceops.index(op) * 2
         changed = tuple(i for i in range(self.shape_len) if resolve(self.sts[reduce_idx].shape[i] != self.sts[reduce_idx + 1].shape[i]))
-        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.UNROLL) if i in changed)
-        grouped_axes = tuple(i for i in self.axes_of(AxisType.GROUP_REDUCE) if i in changed)
+        axes = tuple(i for i in self.axes_of(AxisType.REDUCE, AxisType.GROUP_REDUCE, AxisType.UNROLL) if i in changed)
         if (tc := self.tensor_core) and self.use_tensor_cores == 1:
           # get reduce/upcast axes for the tensor cores
           tc_reduce_axes = self.shape_str_to_axis([f"r{i}" for i in range(len(tc.get_reduce_axes()))])
@@ -486,23 +486,6 @@ class Kernel:
           return ret.replace(src=(tc_uop,), arg=(Ops.ADD, new_axes)) if (new_axes := tuple(i for i in axes if i not in tc_reduce_axes)) else tc_uop
 
         ret = ret.replace(arg = (op.arg[0], axes))
-        if self.group_for_reduces and grouped_axes:
-          local_axes = tuple([i for i,t in enumerate(self.axis_types) if t in (AxisType.LOCAL, AxisType.UPCAST) or i in grouped_axes])
-          slocal, supcast, sgroup = sorted(self.axes_of(AxisType.LOCAL)), sorted(self.axes_of(AxisType.UPCAST)), sorted(grouped_axes)
-          # NOTE: start with UPCAST at the end so it has stride 1 and can merge
-          base_shape = tuple([self.full_shape[i] for i in slocal] + [self.full_shape[i] for i in sgroup] + [self.full_shape[i] for i in supcast])
-          permute_axes = tuple([local_axes.index(i) for i in slocal+sgroup+supcast])
-          local_shape = tuple([s if i in local_axes else 1 for i,s in enumerate(self.full_shape)])
-          local_src_shape = tuple([self.full_shape[i] if i in self.axes_of(AxisType.GLOBAL) else s for i,s in enumerate(local_shape)])
-          st = ShapeTracker.from_shape(base_shape).permute(permute_axes).reshape(local_shape).expand(local_src_shape)
-          local_size = st.real_size()
-          local_buffer = UOp(Ops.DEFINE_LOCAL, op.dtype.ptr(local_size, addrspace=AddrSpace.LOCAL), (), f"temp{self.reduceops.index(op)}")
-          local_load = local_buffer.view(st).load(local_buffer.view(st).store(ret))
-          grouped_reduce = UOp(Ops.REDUCE_AXIS, op.dtype, (local_load,), arg=(op.arg[0], grouped_axes))
-          if op is self.reduceops[-1]: return grouped_reduce
-          st = ShapeTracker.from_shape(tuple([1 if i in grouped_axes else s for i,s in enumerate(local_shape)]))
-          return local_buffer.view(st).load(local_buffer.view(st).store(grouped_reduce))
-
       return ret
     self.finalized = True
     fixed_ast = fixup_ast(self.ast)

tinygrad/codegen/opt/postrange.py

@@ -0,0 +1,197 @@
+import math, itertools
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import UOp, Ops, sint, ssimplify, AxisType, KernelInfo, PatternMatcher, UPat, graph_rewrite
+from tinygrad.helpers import DEBUG, BEAM, getenv
+from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps, KernelOptError
+from tinygrad.renderer import Renderer
+from tinygrad.dtype import dtypes
+from tinygrad.device import Buffer
+
+def flatten_range(r:UOp):
+  off = 2 if r.op is Ops.STORE else 1
+  rngs = r.src[off:]
+  if not len(rngs): return None
+  new_rngs = [x for x in UOp.sink(*rngs).toposort() if x.op is Ops.RANGE]
+  return r.replace(src=r.src[:off]+tuple(new_rngs))
+
+pm_flatten_range = PatternMatcher([
+  # real ranges only
+  (UPat((Ops.REDUCE, Ops.STORE), name="r"), flatten_range),
+])
+
+class RKernel(Kernel):
+  def __init__(self, ast:UOp, opts:Renderer|None=None):
+    self.rng = sorted([u for u in ast.toposort() if u.op is Ops.RANGE and u.vmax > 0], key=lambda x: x.arg)
+    super().__init__(ast, opts)
+    self.sts.clear()
+
+    # convert LOOP to GLOBAL
+    self.replaces = {}
+    if self.opts.has_local:
+      store_rngs = self.ast.src[0].src[2:]
+
+      # filter any not in local stores
+      local_store_rngs = [x.ranges for x in self.ast.toposort() if (x.op is Ops.STORE and x.src[0].dtype.addrspace == AddrSpace.LOCAL) \
+                          or (x.op is Ops.BUFFERIZE and x.arg == AddrSpace.LOCAL)]
+      for ls in local_store_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
+
+      store_rng = [x for x in UOp.sink(*store_rngs).toposort() if x.op is Ops.RANGE] if store_rngs else []
+      rng = [x.replace(arg=(x.arg[0], AxisType.GLOBAL)) if x.arg[1] == AxisType.LOOP and x in store_rng else x for x in self.rng]
+      self.replaces.update(dict(zip(self.rng, rng)))
+      self.rng = rng
+
+      # NOTE: needed for tensor cores
+      self.substitute()
+
+    self.maxarg = max([x.arg[0] for x in self.rng]) if len(self.rng) else 0
+
+  def substitute(self) -> UOp:
+    self.ast = graph_rewrite(self.ast.substitute(self.replaces), pm_flatten_range)
+    self.replaces = {}
+    return self.ast
+
+  def copy(self):
+    self.substitute()
+    return RKernel(self.ast, self.opts)
+
+  # must be done earlier
+  def simplify_merge_adjacent(self): return
+
+  def apply_opt(self, opt:Opt, append_opt:bool=True) -> UOp|None:
+    if opt.op == OptOps.PADTO: raise KernelOptError("PAD is not supported yet. needs INVALID")
+    if opt.op == OptOps.SWAP: raise KernelOptError("SWAP is not supported yet")
+    return super().apply_opt(opt, append_opt)
+
+  def shift_to(self, axis:int, amount:int, new_type:AxisType, top:bool=False, insert_at:int|None=None):
+    old_sz = self.rng[axis].src[0].arg // amount
+    assert old_sz > 0, f"bad old_sz on {axis} {amount} {self.rng[axis]}"
+
+    self.maxarg += 1
+    new_rng = UOp.range(amount, self.maxarg, new_type)
+
+    if old_sz == 1:
+      self.replaces[self.rng[axis]] = new_rng
+      self.rng.insert(insert_at if insert_at is not None else len(self.rng), new_rng)
+      del self.rng[axis]
+    else:
+      replaced_rng = self.rng[axis].replace(src=(UOp.const(dtypes.int, old_sz),))
+      self.replaces[self.rng[axis]] = (new_rng * old_sz + replaced_rng) if top else (replaced_rng * amount + new_rng)
+      self.rng[axis] = replaced_rng
+      self.rng.insert(insert_at if insert_at is not None else len(self.rng), new_rng)
+    return new_rng
+
+  @property
+  def axis_types(self) -> list[AxisType]: return [x.arg[-1] for x in self.rng]
+  @property
+  def shape_len(self): return len(self.rng)
+
+  @property
+  def full_shape(self) -> tuple[sint, ...]: return tuple([ssimplify(x.src[0]) for x in self.rng])
+  @property
+  def output_shape(self) -> tuple[sint, ...]:
+    if self.ast.src[0].op is not Ops.STORE: return ()
+    return tuple([ssimplify(x.src[0]) for x in self.ast.src[0].src[2:]])
+
+  def get_optimized_ast(self, name_override:str|None=None) -> UOp:
+    ret = self.substitute()
+    kernel_name = ret.arg.name if ret.arg is not None and ret.arg.name != "test" else self.name if name_override is None else name_override
+    rarg = KernelInfo(kernel_name, tuple(self.axis_types), self.dont_use_locals, tuple(self.applied_opts))
+    return ret.replace(arg=rarg)
+
+  # does nothing
+  @axis_types.setter
+  def axis_types(self, value): pass
+
+  def _apply_tc_opt(self, use_tensor_cores:int, axis:int, tc_select:int, opt_level:int) -> bool:
+    reduceops = [x for x in self.ast.toposort() if x.op is Ops.REDUCE]
+    if not len(reduceops): raise KernelOptError("no reduce ops for TensorCore")
+    reduceop = reduceops[0]
+    if use_tensor_cores and reduceop is not None and reduceop.arg is Ops.ADD:
+      mul = reduceop.src[0] if reduceop.src[0].op is not Ops.CAST else reduceop.src[0].src[0]
+      if mul.op is not Ops.MUL: return False
+      in0, in1 = mul.src
+      tensor_cores = self.opts.tensor_cores if tc_select == -1 else [self.opts.tensor_cores[tc_select]]
+      for tc in tensor_cores:
+        if tc.dtype_in == in0.dtype.scalar() and tc.dtype_in == in1.dtype.scalar() and tc.dtype_out == reduceop.dtype.scalar():
+          # early realize for TC
+          self.substitute()
+
+          # tensor cores have three ranges. X, Y, and REDUCE
+          in0_ranges = sorted([u for u in in0.ranges if u not in in1.ranges], key=lambda x: x.arg[0])
+          in1_ranges = sorted([u for u in in1.ranges if u not in in0.ranges], key=lambda x: x.arg[0])
+          red_ranges = sorted(reduceop.src[1:], key=lambda x: x.arg[0])
+          if DEBUG >= 3:
+            print(f"TC({axis}): {[(x.arg[0],x.vmax+1) for x in in0_ranges]}",
+                              f"{[(x.arg[0],x.vmax+1) for x in in1_ranges]} {[(x.arg[0],x.vmax+1) for x in red_ranges]}")
+          if not len(in0_ranges) or not len(in1_ranges) or not len(red_ranges): return None
+
+          # pick ranges
+          # NOTE: why are in1 and in0 switched?
+          axis_choices = list(itertools.product(in1_ranges, in0_ranges, red_ranges))
+          if not (axis < len(axis_choices)): return None
+          axes = axis_choices[axis]
+
+          # do optimizations and save the ranges
+          try:
+            for i,a in enumerate(axes):
+              if a.src[0].divides(tc.dims[i]) is None:
+                self.apply_opt(Opt(OptOps.PADTO, self.rng.index(a), tc.dims[i]), append_opt=False) # PADTO might fail
+          except KernelOptError: continue
+          ne: list[UOp] = []
+          for opt in tc.opts:
+            ne.append(self.apply_opt(Opt({"u":OptOps.UPCAST, "l":OptOps.LOCAL}[opt[0]], axes[int(opt[1])].arg[0], 2), append_opt=False))
+          reduce_axis = [self.rng[i] for i in self.axes_of(AxisType.REDUCE)].index(axes[2])
+          for _, amt in tc.get_reduce_axes():
+            ne.append(self.apply_opt(Opt(OptOps.UNROLL, reduce_axis, amt), append_opt=False))
+
+          if use_tensor_cores != 2:
+            # fix the srcs
+            reduceop = [x for x in self.ast.toposort() if x.op is Ops.REDUCE][0]
+            tne = [x.replace(tag=1) for x in ne]
+            ret = reduceop.substitute(dict(zip(ne, tne)))
+            srcs = list((ret.src[0] if ret.src[0].op is not Ops.CAST else ret.src[0].src[0]).src)
+            srcs = [x.substitute(dict(zip(tne, [ne[i] for i in p]))) for x,p in zip(srcs, tc.permutes_for_shape_str(tc.base_shape_str()))]
+
+            # get reduce/upcast axes for the tensor cores
+            tc_reduce_axes = self.shape_str_to_axis([f"r{i}" for i in range(len(tc.get_reduce_axes()))])
+            base_upcast_axes = tuple([(s,2) for s in self.shape_str_to_axis(tc.base_upcast_axes())])
+            tc_upcast_axes = tuple([base_upcast_axes[:int(math.log2(tc.elements_per_thread[i]))] for i in range(3)])
+
+            # axes to range number (was done in lowerer)
+            tc_upcast_axes = tuple([tuple([(self.rng[a].arg[0], sz) for a,sz in v]) for v in tc_upcast_axes])
+            tc_reduce_axes = tuple([self.rng[a].arg[0] for a in tc_reduce_axes])
+
+            # construct the op
+            # TODO: remove tc_upcast_axes from the arg
+            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.opts.device, tc.threads, tc_upcast_axes, tc_reduce_axes)
+            wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(tc.elements_per_thread[2]), src=(
+              UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(tc.elements_per_thread[0]), src=(srcs[0],), arg=tc_upcast_axes[0], tag=1),
+              UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(tc.elements_per_thread[1]), src=(srcs[1],), arg=tc_upcast_axes[1], tag=1),
+              UOp.const(tc.dtype_out.vec(tc.elements_per_thread[2]), 0.0)), arg=wmma_arg, tag=1)
+            tc_uop = UOp(Ops.UNROLL, tc.dtype_out, (wmma,), arg=tc_upcast_axes[2], tag=1)
+
+            # preserve extra reduces
+            reduce_ranges = [x for x in UOp.sink(*reduceop.src[1:]).toposort() if x.op is Ops.RANGE and x.arg[0] not in tc_reduce_axes]
+            if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, tc_uop.dtype, (tc_uop,)+tuple(reduce_ranges), Ops.ADD)
+            self.ast = self.ast.substitute({reduceop: tc_uop})
+          return True
+    return False
+
+def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
+  glbls = sorted([x for x in ast.parents if x.op is Ops.DEFINE_GLOBAL], key=lambda x: x.arg)
+  return [Buffer(dname, x.dtype.size, x.dtype.base) for x in glbls]
+
+def apply_ropt(ast:UOp, renderer:Renderer):
+  k = RKernel(ast, opts=renderer)
+  if BEAM >= 1:
+    from tinygrad.codegen.opt.search import beam_search
+    kb = RKernel(ast, opts=renderer)
+    rawbufs = bufs_from_ast(ast, renderer.device)
+    k = beam_search(kb, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
+  elif ast.arg is not None: k.apply_opts(ast.arg.opts_to_apply)
+  return k.get_optimized_ast()
+
+pm_postrange_opt = pm_flatten_range+PatternMatcher([
+  (UPat(Ops.SINK, name="ast"), lambda ctx,ast: apply_ropt(ast, ctx) if ast.arg is None or \
+   (ast.arg is not None and ast.arg.opts_to_apply is not None) else None),
+])

tinygrad/codegen/opt/search.py

@@ -1,5 +1,5 @@
-from typing import cast, Callable
-import itertools, functools, random, math, time, multiprocessing, traceback, signal, atexit
+from typing import cast
+import functools, math, time, multiprocessing, traceback, signal, atexit
 from collections import defaultdict
 from dataclasses import replace
 from tinygrad.uop.ops import UOp, Ops, Variable, sym_infer, AxisType
@@ -201,15 +201,3 @@ def beam_search(lin:Kernel, rawbufs:list[Buffer], amt:int, allow_test_size=True,
   if CACHELEVEL >= 1: diskcache_put("beam_search", key, beam[0][0].applied_opts)
   if BEAM_DEBUG: print(f"BEAM_SEARCH: final tm={time_to_str(beam[0][1], w=0)}, applied_opts={beam[0][0].applied_opts}")
   return beam[0][0]
-
-def optimize_local_size(_prg:Callable, global_size:list[int], rawbufs:list[Buffer]) -> list[int]:
-  test_rawbuffers = [Buffer(rawbufs[0].device, rawbufs[0].size, rawbufs[0].dtype).allocate(), *rawbufs[1:]] if rawbufs[0] in rawbufs[1:] else rawbufs
-  MAX_WORKGROUP = 1024
-  local_dims = [[x for x in set([sz, 1, 2, 4, 8, 16, 32, 64, 128, 256, MAX_WORKGROUP]) if x<=sz] for sz in global_size]
-  local_sizes = [list(x) for x in itertools.product(*local_dims) if prod(x) <= MAX_WORKGROUP] * 2  # try each valid size twice
-  def try_exec(local_size):
-    try: return _prg(*[x._buf for x in test_rawbuffers], global_size=[g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)], local_size=local_size, wait=True)  # noqa: E501
-    except Exception: return float('inf')
-  ret = min([(try_exec(local_size), local_size) for local_size in random.sample(local_sizes, len(local_sizes))])
-  assert not math.isinf(ret[0]), "all optimize_local_size exec failed"
-  return ret[1]

tinygrad/codegen/opt/swizzler.py

@@ -128,7 +128,8 @@ fix_kernel_ops = view_left_through_load+PatternMatcher([
   (UPat(Ops.VIEW, src=(UPat.cvar(),), name="self"),
    lambda self: UOp.where(UOp(Ops.VALID, dtypes.bool, (UOp(Ops.VIEW, arg=self.st),)), self.const_like(self.base.arg), 0)),
   # no ImageDType after index
-  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW}, name="x"), lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
+  (UPat(GroupOp.All-{Ops.DEFINE_GLOBAL, Ops.VIEW, Ops.INDEX}, name="x"),
+   lambda x: x.replace(dtype=x.dtype.base) if isinstance(x.dtype, ImageDType) else None),
   # if this kernel also assigns to the loaded buffer, ensure we can index it correctly
   (UPat(Ops.LOAD, src=(UPat.var("glbl").view(name="view"),)), check_load_st),
 ])

tinygrad/codegen/opt/tc.py

@@ -22,6 +22,15 @@ class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x
   def permutes_for_shape_str(self, shape_str:list[str]) -> tuple[tuple[int, ...], tuple[int, ...]]:
     ret = [[shape_str.index(remap[ss]) if ss in remap else i for i,ss in enumerate(shape_str)] for remap in self._remaps()]
     return tuple(ret[0]), tuple(ret[1])
+  @functools.cache  # pylint: disable=method-cache-max-size-none
+  def base_shape_str(self) -> list[str]:
+    ret = []
+    cnt = {'u': 0, 'l': 0}
+    for opt in self.opts:
+      ret.append(f"{opt[0]}{cnt[opt[0]]}")
+      cnt[opt[0]] += 1
+    # assumes you do the UNROLL after the opts
+    return ret + [f"r{i}" for i in range(len(self.get_reduce_axes()))]
   def get_reduce_axes(self): return [(i, 2) for i in range(int(math.log2(self.dims[2])))]
   def get_upcast_axes(self): return [opt for opt in self.opts if opt[0] == "u"]
   def get_local_axes(self): return [opt for opt in self.opts if opt[0] == "l"]

tinygrad/device.py

@@ -304,7 +304,7 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device == "METAL": return not CI
     if device in {"CUDA", "NV"}: return not CI and not getenv("PTX")
     if device in {"CPU", "LLVM"}: return not CI and platform.machine() in {"arm", "arm64", "aarch64", "x86_64", "amd64"}
-    return device == "AMD"
+    return device in {"AMD", "PYTHON"}
   if dtype in dtypes.fp8s:
     # not supported yet - in progress
     return False

tinygrad/dtype.py

@@ -108,7 +108,6 @@ class dtypes:
     if isinstance(val, tuple):
       assert len(val) == dtype.count, f"mismatch {val} {dtype}"
       return tuple(dtypes.as_const(x, dtype) for x in val)
-    # TODO: should truncate here
     return int(val) if dtypes.is_int(dtype) else float(val) if dtypes.is_float(dtype) else bool(val)
   @staticmethod
   @functools.cache
@@ -215,15 +214,14 @@ def sum_acc_dtype(dt:DType):
   return least_upper_dtype(dt, to_dtype(getenv("SUM_DTYPE", "float32")))
 
 def truncate_fp16(x):
-  try: return struct.unpack("@e", struct.pack("@e", float(x)))[0]
+  try: return struct.unpack('e', struct.pack('e', float(x)))[0]
   except OverflowError: return math.copysign(math.inf, x)
 
-def truncate_bf16(x):
-  max_bf16 = struct.unpack('f', struct.pack('I', 0x7f7f0000))[0]
-  if abs(x) > max_bf16: return math.copysign(math.inf, x)
-  f32_int = struct.unpack('I', struct.pack('f', x))[0]
-  bf = struct.unpack('f', struct.pack('I', f32_int & 0xFFFF0000))[0]
-  return bf
+def float_to_bf16(x):
+  if not math.isfinite(x): return x
+  u = struct.unpack('I', struct.pack('f', x))[0]
+  u = (u + 0x7FFF + ((u >> 16) & 1)) & 0xFFFF0000
+  return struct.unpack('f', struct.pack('I', u))[0]
 
 # fp8-float conversions based on https://gitlab.com/nvidia/headers/cuda-individual/cudart/-/blob/main/cuda_fp8.hpp
 def float_to_fp8(x: float, dtype: DType) -> int:
@@ -288,7 +286,7 @@ def fp8_to_float(x: int, dtype: DType) -> float:
   return float(float32_val)
 
 truncate: dict[DType, Callable] = {dtypes.bool: bool,
-  dtypes.float16: truncate_fp16, dtypes.bfloat16: truncate_bf16,
+  dtypes.float16: truncate_fp16, dtypes.bfloat16: lambda x: float_to_bf16(float(x)),
   **{fp8: (lambda x, dtype=fp8: fp8_to_float(float_to_fp8(x, dtype), dtype)) for fp8 in dtypes.fp8s},
   dtypes.float32: lambda x: ctypes.c_float(x).value, dtypes.float64: lambda x: ctypes.c_double(x).value,
   dtypes.uint8: lambda x: ctypes.c_uint8(x).value, dtypes.uint16: lambda x: ctypes.c_uint16(x).value,
@@ -300,6 +298,7 @@ truncate: dict[DType, Callable] = {dtypes.bool: bool,
 
 def _to_np_dtype(dtype:DType) -> type|None:
   import numpy as np
+  if dtype == dtypes.bfloat16: return np.float32
   return np.dtype(dtype.fmt).type if dtype.fmt is not None else None
 def _from_np_dtype(npdtype:'np.dtype') -> DType: # type: ignore [name-defined] # noqa: F821
   import numpy as np
@@ -308,6 +307,8 @@ def _from_np_dtype(npdtype:'np.dtype') -> DType: # type: ignore [name-defined] #
 @functools.cache
 def _to_torch_dtype(dtype:DType) -> 'torch.dtype'|None:  # type: ignore [name-defined] # noqa: F821
   import numpy as np, torch
+  if dtype == dtypes.uint64: return torch.uint64
+  if dtype == dtypes.bfloat16: return torch.bfloat16
   # NOTE: torch doesn't expose this mapping with a stable API
   try: return torch.from_numpy(np.array([], dtype=_to_np_dtype(dtype))).dtype
   except TypeError: return None

tinygrad/engine/memory.py

@@ -23,12 +23,13 @@ def _internal_memory_planner(buffers:list[list[Buffer]], noopt_buffers=None, ign
   # Sort buffer operations in timeline order. Two events: buffer is allocated or buffer is freed.
   buffer_requests = sorted([((first_appearance[buf], True), buf) for buf in first_appearance.keys()] + \
                            [((last_appearance[buf] + 1, False), buf) for buf in first_appearance.keys()], key=lambda x: x[0])
+  total_memory = sum(round_up(buf.nbytes, min_block_size:=0x1000) for buf in first_appearance.keys()) * 2 # *2 for fragmentation (which is about 15%)
 
   # Try to suballocate from a shared buffer managed by global_planner using TLSFAllocator.
   # Also track buffer replacements for buffers that do not support suballocation.
   buffer_replace:dict[Buffer, tuple[Buffer|None, int|None]] = {}
   reuse_buffers:dict[tuple, list[Buffer]] = defaultdict(list)
-  global_planner:dict[str, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(1 << 44, block_size=0x1000, lv2_cnt=32)))
+  global_planner:dict[str, tuple[int, TLSFAllocator]] = defaultdict(lambda: (0, TLSFAllocator(total_memory, block_size=min_block_size, lv2_cnt=32)))
   for (_, is_open_ev), buf in buffer_requests:
     # Check if suballocation is possible for the given buffer and device.
     if hasattr(Device[buf.device].allocator, "_offset") and not isinstance(buf.dtype, ImageDType):

tinygrad/engine/realize.py

@@ -1,8 +1,8 @@
-from typing import cast, Generator
-import time, pprint, decimal
+from typing import cast, Generator, Callable
+import time, pprint, decimal, random, itertools, math
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import all_same, colored, DEBUG, GlobalCounters, ansilen, BEAM, NOOPT, all_int, CAPTURING, Metadata, TRACEMETA, TracingKey
-from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events
+from tinygrad.helpers import DEVECTORIZE, time_to_str, VALIDATE_WITH_CPU, getenv, cpu_profile, PROFILE, ProfilePointEvent, cpu_events, prod
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, Variable, sym_infer, graph_rewrite, print_uops, track_rewrites, KernelInfo
 from tinygrad.device import Device, Buffer
 from tinygrad.renderer import Renderer, ProgramSpec, Estimates
@@ -59,6 +59,20 @@ class Runner:
   def __call__(self, rawbufs:list[Buffer], var_vals:dict[Variable, int], wait=False) -> float|None:
     raise NotImplementedError("override this")
 
+def optimize_local_size(_prg:Callable, global_size:list[int], rawbufs:list[Buffer]) -> list[int]:
+  test_rawbuffers = [Buffer(rawbufs[0].device, rawbufs[0].size, rawbufs[0].dtype).allocate(), *rawbufs[1:]] if rawbufs[0] in rawbufs[1:] else rawbufs
+  MAX_WORKGROUP = 1024
+  local_dims = [[x for x in set([sz, 1, 2, 4, 8, 16, 32, 64, 128, 256, MAX_WORKGROUP]) if x<=sz] for sz in global_size]
+  local_sizes = [list(x) for x in itertools.product(*local_dims) if prod(x) <= MAX_WORKGROUP] * 2  # try each valid size twice
+  def try_exec(local_size):
+    try:
+      return _prg(*[x._buf for x in test_rawbuffers],global_size=[g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)],
+                  local_size=local_size, wait=True)
+    except Exception: return float('inf')
+  ret = min([(try_exec(local_size), local_size) for local_size in random.sample(local_sizes, len(local_sizes))])
+  assert not math.isinf(ret[0]), "all optimize_local_size exec failed"
+  return ret[1]
+
 class CompiledRunner(Runner):
   def __init__(self, p:ProgramSpec, precompiled:bytes|None=None, prg=None):
     if DEBUG >= 4: print(p.src)
@@ -76,8 +90,6 @@ class CompiledRunner(Runner):
   def __call__(self, rawbufs:list[Buffer], var_vals:dict[Variable, int], wait=False) -> float|None:
     global_size, local_size = self.p.launch_dims(var_vals)
     if global_size is not None and local_size is None and all_int(self.p.global_size): # type: ignore[arg-type]
-      # TODO: this is copied from get_program
-      from tinygrad.codegen.opt.search import optimize_local_size
       local_size = optimize_local_size(self._prg, global_size, rawbufs)
       global_size = [g//l if g%l == 0 else g/l for g,l in zip(global_size, local_size)]
       self.p = replace(self.p, global_size=global_size, local_size=local_size)
@@ -160,10 +172,15 @@ class ExecItem:
       if DEBUG >= 2:
         lds_est = sym_infer(self.prg.estimates.lds, var_vals)
         mem_est = min(mem_est, lds_est)   # there can't be more memory accessed than loads/stores. remove this when symbolic is fixed
+        header_color = 'magenta' if jit else ('green' if self.prg.first_run else None)
         ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
-        print(f"{colored(f'*** {self.prg.device[:7]:7s} {GlobalCounters.kernel_count:4d}', 'magenta' if jit else ('green' if self.prg.first_run else None))} {self.prg.display_name+' '*(44-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:5.2f} GB " +  # noqa: E501
-              (str() if et is None else f"tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({op_est/((et or 1e-20)*1e9):9.2f} GFLOPS {mem_est/((et or 1e-20)*1e9):6.1f}|{lds_est/((et or 1e-20)*1e9):<7.1f} GB/s)" +  # noqa: E501
-               f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in self.metadata] if self.metadata else ''}"))
+        flops, membw, ldsbw = op_est/(et or 1e-20), mem_est/(et or 1e-20), lds_est/(et or 1e-20)
+        flops_str = f"{flops*1e-9:9.2f} GFLOPS" if flops < 1e14 else colored(f"{flops*1e-12:9.2f} TFLOPS", 'green')
+        mem_str = f"{membw*1e-9:6.1f}|{ldsbw*1e-9:<7.1f} GB/s" if membw < 1e13 else colored(f"{membw*1e-12:6.1f}|{ldsbw*1e-12:<7.1f} TB/s", 'green')
+        print(f"{colored(f'*** {self.prg.device[:7]:7s} {GlobalCounters.kernel_count:4d}', header_color)}"+
+          f" {self.prg.display_name+' '*(44-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:5.2f} GB"+
+          ("" if et is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
+          f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in self.metadata] if self.metadata else ''}")
       self.prg.first_run = False
     return et
 

tinygrad/frontend/onnx.py

@@ -21,9 +21,9 @@ class AttributeType(enum.IntEnum):
   ONNX attribute type identifiers.
   Reference: https://github.com/onnx/onnx/blob/rel-1.18.0/onnx/onnx.proto3#L128-L145
   """
-  FLOAT = 1; INT = 2; STRING = 3; TENSOR = 4; FLOATS = 6; INTS = 7; STRINGS = 8 # noqa: E702
+  FLOAT = 1; INT = 2; STRING = 3; TENSOR = 4; GRAPH = 5; FLOATS = 6; INTS = 7; STRINGS = 8 # noqa: E702
 
-  def to_field_name(self) -> str: return {1: "f", 2: "i", 3: "s", 4: "t", 6: "floats", 7: "ints", 8: "strings"}[self.value]
+  def to_field_name(self) -> str: return {1: "f", 2: "i", 3: "s", 4: "t", 5: "g", 6: "floats", 7: "ints", 8: "strings"}[self.value]
 
 class OnnxDataType(enum.IntEnum):
   """
@@ -266,6 +266,7 @@ class OnnxPBParser:
         case 3: obj["i"] = self.reader.read_int64()
         case 4: obj["s"] = self.reader.read_bytes().data().tobytes().decode("utf8")
         case 5: obj["t"] = self._parse_TensorProto()['parsed_tensor']
+        case 6: obj["g"] = OnnxRunner._from_subgraph(self._parse_GraphProto())
         case 7: obj["floats"].append(self.reader.read_float())
         case 8: obj["ints"].append(self.reader.read_int64())
         case 9: obj["strings"].append(self.reader.read_bytes().data().tobytes().decode("utf8"))
@@ -401,8 +402,11 @@ class OnnxRunner:
   """
   def __init__(self, model_path: Tensor | str | pathlib.Path):
     model = OnnxPBParser(model_path, load_external_data=True).parse()
-    graph = model["graph"]
+    self._init_from_graph(model["graph"])
+
+  def _init_from_graph(self, graph: dict, is_subgraph: bool = False):
     self.is_training = any(n['parsed_node'].opset_id.domain in {Domain.AI_ONNX_TRAINING, Domain.AI_ONNX_PREVIEW_TRAINING} for n in graph["node"])
+    self.graph_name = graph["name"] if is_subgraph else ""
     self.graph_values = {"": None, **{i["name"]: i["parsed_tensor"] for i in graph["initializer"]}}
     self.graph_inputs = {i["name"]: i["parsed_type"] for i in graph["input"] if i["name"] not in self.graph_values}
     self.graph_outputs = tuple(o["name"] for o in graph["output"])
@@ -414,6 +418,12 @@ class OnnxRunner:
     self.variable_dims: dict[str, int] = {}
     self.onnx_ops = onnx_ops
 
+  @classmethod
+  def _from_subgraph(cls, graph: dict) -> "OnnxRunner":
+    subgraph = cls.__new__(cls)
+    subgraph._init_from_graph(graph, is_subgraph=True)
+    return subgraph
+
   def _parse_input(self, name: str, value: Any, spec: OnnxValue):
     if spec.is_optional and value is None: return None
     if spec.is_sequence:
@@ -445,9 +455,10 @@ class OnnxRunner:
     return {name:Tensor.empty(*spec.shape, device=device, dtype=dtype or spec.dtype) for name, spec in self.graph_inputs.items()}
 
   def to(self, device:str|None):
-    self.graph_values = {k:v.to(device) if isinstance(v, Tensor) else v for k,v in self.graph_values.items()}
+    self.graph_values = {k: (v.to(device) if isinstance(v, Tensor) else v) for k,v in self.graph_values.items()}
     self.graph_nodes = tuple(OnnxNode(n.op, n.opset_id, tuple(n.inputs), tuple(n.outputs),
-                                      {k:v.to(device) if isinstance(v, Tensor) else v for k,v in n.opts.items()}) for n in self.graph_nodes)
+                                      {k: (v.to(device) if isinstance(v, (Tensor, OnnxRunner)) else v) for k,v in n.opts.items()})
+                                      for n in self.graph_nodes)
     return self
 
   def __call__(self, inputs:dict[str, Any], debug=debug):
@@ -461,9 +472,9 @@ class OnnxRunner:
 
       # provide additional opts
       if node.op == "Split" and 'num_outputs' not in opts: opts['num_outputs'] = len(node.outputs)
-      if node.op == "Gradient": opts['intermediate_tensors'] = self.graph_values
+      if node.op in {"Gradient", "If"}: opts['intermediate_tensors'] = self.graph_values
 
-      if debug >= 1: print(f"{num}: op '{node.op}' opt {opts}")
+      if debug >= 1: print((f"[{self.graph_name}] " if self.graph_name else "") + f"{num}: op '{node.op}' opt {opts}")
       if debug >= 2 and node.inputs: print("\tinputs:\n" + "\n".join(f"\t\t{x} - {i!r}" for x,i in zip(node.inputs, inps)))
       ret = self._select_op(node.op, node.opset_id)(*inps, **opts)
       ret = ret if isinstance(ret, tuple) else (ret,)
@@ -543,6 +554,23 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     return __decorator
 
   # ***** Property/Graph Ops *****
+  def If(condition:Tensor, else_branch:OnnxRunner, then_branch:OnnxRunner, intermediate_tensors:dict[str, Tensor]):
+    def run_branch(branch:OnnxRunner):
+      branch.graph_values.update(intermediate_tensors)
+      out = branch({k:intermediate_tensors[k] for k in branch.graph_inputs.keys()})
+      # dereference intermediate tensors so Buffer can be deallocated
+      for k in intermediate_tensors: del branch.graph_values[k]
+      return out
+    # both branch must be ran before the condition can be evaluated
+    else_out, then_out = run_branch(else_branch), run_branch(then_branch)
+    assert len(else_out) == len(then_out), f"else_out and then_out must have the same number of outputs: {len(else_out)} != {len(then_out)}"
+    # can use where op when output shape is the same
+    if all(t.shape == e.shape for t,e in zip(then_out.values(), else_out.values())):
+      return tuple(condition.where(t,e) for t,e in zip(then_out.values(), else_out.values()))
+    # otherwise, use condition to select the output in python
+    cond = _resolve_const(_cached_to_python_const(condition))
+    return tuple(t if cond else e for t,e in zip(then_out.values(), else_out.values()))
+
   def Identity(x:Tensor): return x
   def Constant(sparse_value:Tensor|None=None, value:Tensor|None=None, value_float:float|None=None, value_floats:list[float]|None=None,
               value_int:int|None=None, value_ints:list[int]|None=None, value_string:str|None=None, value_strings:list[str]|None=None):

tinygrad/gradient.py

@@ -22,11 +22,10 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.SQRT, name="ret"), lambda ctx, ret: (ctx / (ret*2),)),
   (UPat((Ops.CMPLT, Ops.CMPNE)), lambda: (None, None)),
   (UPat(Ops.ADD), lambda ctx: (ctx, ctx)),
-  (UPat(Ops.POW, name="ret"), lambda ctx, ret:
-   (ctx*(ret.src[0].eq(0) & ret.src[1].eq(0)).where(ret.src[1], ret.src[1]*ret.src[0].pow(ret.src[1]-1)),
-    ctx*ret.src[0].eq(0).where((ret.src[1]<0).where(ret.const_like(-math.inf), ret.const_like(0)), ret*ret.src[0].log2()*math.log(2.0)))),
-  (UPat(Ops.MAX, name="ret"), lambda ctx, ret: ((ret.src[0]>ret.src[1]).where(ctx, (ret.src[0]!=ret.src[1]).where(ctx.const_like(0), ctx * 0.5)),
-                                                (ret.src[0]<ret.src[1]).where(ctx, (ret.src[0]!=ret.src[1]).where(ctx.const_like(0), ctx * 0.5)))),
+  (UPat(Ops.POW, name="ret", src=(UPat.var("b"), UPat.var("e"))), lambda ctx, ret, b, e:
+    (ctx * (b.eq(0)&e.eq(0)).where(e, e*b.pow(e-1)), ctx * b.eq(0).where((e<0).where(ret.const_like(-math.inf), 0), ret*b.log2()*math.log(2.0)))),
+  (UPat(Ops.MAX, name="ret", src=(UPat.var("x"), UPat.var("y"))), lambda ctx, ret, x, y:
+    ((x>y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)), (x<y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)))),
   (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
   (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),
   (UPat(Ops.REDUCE_AXIS, name="ret"), reduce_gradient),

tinygrad/helpers.py

@@ -140,7 +140,7 @@ DONT_REALIZE_EXPAND, DONT_GROUP_REDUCES = ContextVar("DONT_REALIZE_EXPAND", 0),
 QUANTIZE, VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("QUANTIZE", 0), ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
 CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)
 ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, AMD_LLVM = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0), ContextVar("AMD_LLVM", 1)
-RANGEIFY = ContextVar("RANGEIFY", 0)
+RANGEIFY, POSTOPT, FUSE_ATTENTION = ContextVar("RANGEIFY", 0), ContextVar("POSTOPT", 1), ContextVar("FUSE_ATTENTION", 0)
 
 @dataclass(frozen=True)
 class Metadata:
@@ -196,7 +196,7 @@ class Profiling(contextlib.ContextDecorator):
 @dataclass(frozen=True)
 class TracingKey:
   display_name:str                       # display name of this trace event
-  keys:tuple[str, ...]=()                # optional keys to search for related traces
+  keys:tuple[Any, ...]=()                # optional keys to search for related traces
   cat:str|None=None                      # optional category to color this by
   ret:Any=None
 

tinygrad/nn/state.py

@@ -274,9 +274,9 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
   Converts ggml tensor data to a tinygrad tensor.
 
   Supported native types: float32 (id: 0), float16 (id: 1), int8 (id: 16), int16 (id: 17), int32 (id: 18)
-  Supported quantized types: Q4_0 (id: 2), Q4_1 (id: 3), Q8_0 (id: 8), Q6_K (id: 14)
+  Supported quantized types: Q4_0 (id: 2), Q4_1 (id: 3), Q8_0 (id: 8), Q6_K (id: 14), MXFP4 (id: 39)
   """
-  # https://github.com/ggerganov/ggml/blob/6dccc647264f5429df2624f36138f601e7ce23e5/include/ggml.h#L356
+  # https://github.com/ggerganov/ggml/blob/323951f1bdcdfbd5b5ff3a9a7c3770e63b1a560e/include/ggml.h#L356
 
   # native types
   if (dtype := { 0: dtypes.float32, 1: dtypes.float16, 16: dtypes.int8, 17: dtypes.int16, 18: dtypes.int32 }.get(ggml_type)) is not None:
@@ -288,7 +288,7 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
     return t.unsqueeze(-1).expand((*t.shape,8//b)).idiv(shift_tensor).bitwise_and(bitmask).transpose(-1, -2).flatten(-2)
 
   # map to (number of elements, number of bytes)
-  if (nelements_nbytes := { 2: (32, 18), 3: (32, 20), 14: (256, 210), 8: (32, 34) }.get(ggml_type)) is not None:
+  if (nelements_nbytes := { 2: (32, 18), 3: (32, 20), 14: (256, 210), 8: (32, 34), 39: (32, 17) }.get(ggml_type)) is not None:
     blocks = t[:(n//nelements_nbytes[0])*nelements_nbytes[1]].reshape((-1, nelements_nbytes[1]))
     if ggml_type == 2: return (q_to_uint8(blocks[:,2:], 4).bitcast(dtypes.int8) - 8) * blocks[:,:2].bitcast(dtypes.float16).cast(dtypes.float32)
     if ggml_type == 3:
@@ -300,6 +300,17 @@ def ggml_data_to_tensor(t: Tensor, n: int, ggml_type: int) -> Tensor:
       scales = blocks[:,192:208].bitcast(dtypes.int8).unsqueeze(-1).expand((-1, 16, 16)).reshape((-1, 256))
       d = blocks[:,-2:].bitcast(dtypes.float16).cast(dtypes.float32).expand((-1, 256))
       return d * (xl.bitwise_or(xh).bitcast(dtypes.int8) - 32).flatten(-2) * scales
+    if ggml_type == 39:
+      e_int = blocks[:, 0].cast(dtypes.int32)
+      d = ((e_int >= 2).cast(dtypes.float32) * (e_int.cast(dtypes.float32) - 128).exp2() +
+           (e_int == 1).cast(dtypes.float32) * 2.0**(-127) +
+           (e_int == 0).cast(dtypes.float32) * 2.0**(-128)).unsqueeze(-1)
+      codes = q_to_uint8(blocks[:, 1:17], 4)
+      sign = 1.0 - codes.rshift(3).cast(dtypes.float32) * 2.0
+      exp, mant = codes.rshift(1).bitwise_and(0x3).cast(dtypes.float32), codes.bitwise_and(0x1).cast(dtypes.float32)
+      fp4_val = sign * ((exp != 0).cast(dtypes.float32) * (1.0 + 0.5 * mant) * (exp - 1.0).exp2() +
+                        (exp == 0).cast(dtypes.float32) * 0.5 * mant)
+      return (fp4_val * d).flatten(-2)[:n]
   raise ValueError(f"GGML type '{ggml_type}' is not supported!")
 
 @accept_filename

tinygrad/renderer/cstyle.py

@@ -157,7 +157,7 @@ class CStyleLanguage(Renderer):
       # naming
       prefix = None
       if u.op is Ops.SPECIAL: r[u] = u.arg[0]
-      elif u.op is Ops.RANGE: r[u] = f"ridx{u.arg[0]}" if u.arg[0] >= 0 else f"ridxm{-u.arg[0]}"
+      elif u.op is Ops.RANGE: r[u] = "ridx"+'_'.join([str(x) if x >= 0 else "m"+str(-x) for x in u.arg[0:-1]])
       else:
         prefix = {Ops.WMMA: "wmma", Ops.DEFINE_LOCAL: "temp", Ops.CONST: "const",
                   Ops.CAST: "cast", Ops.BITCAST: "cast", Ops.GEP: "gep", Ops.VECTORIZE: "cast", Ops.PRECAST: "precast",

tinygrad/renderer/ptx.py

@@ -119,7 +119,7 @@ string_rewrite = PatternMatcher([
     ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[x.src[0]], ctx.r[src0.src[0]], dtypes.int, ctx.types[dtypes.int]),
     f"@{ctx.r[x]} bra LOOP_{ctx.r[src0][1:]};"]),
   (UPat(Ops.DEFINE_LOCAL, name="x"),
-   lambda ctx, x: [f".shared .align 16 .b8 {x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, {x.arg}[0];"]),
+   lambda ctx, x: [f".shared .align 16 .b8 local{x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, local{x.arg}[0];"]),
   (UPat(Ops.IF, name="x"), lambda ctx, x: f"@!{ctx.r[x.src[0]]} bra IF_{ctx.r[x.src[0]][1:]}_{ctx.uops.index(x)};"),
   (UPat(Ops.ENDIF, name="x"), lambda ctx, x: f"IF_{ctx.r[x.src[0].src[0]][1:]}_{ctx.uops.index(x.src[0])}:"),
   (UPat(Ops.WMMA, name="x"), lambda ctx, x: list(render_wmma(ctx, x))),
@@ -215,7 +215,7 @@ class PTXRenderer(Renderer):
                        [ssa("wmma_acc", dtype="b32") for _ in range(0, len(r[u.src[2]]), 4 // u.dtype.scalar().itemsize)]]
         r[u] = [ssa("wmma", dtype=self.types[u.dtype.scalar()]) for _ in range(u.dtype.count)]
       prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.ENDRANGE: ("pred", "pred"), Ops.RANGE: ("ridx", None),
-        Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL:("local",self.types[dtypes.ulong]),
+        Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL: ("local",self.types[dtypes.ulong]),
         Ops.DEFINE_GLOBAL: ("dat", self.types[dtypes.ulong]), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
       if prefix: r[u] = ssa(prefix, u, dtype)
 

tinygrad/runtime/ops_amd.py

@@ -7,8 +7,7 @@ from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, H
 from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator
 from tinygrad.uop.ops import sint
 from tinygrad.device import Compiled, DMAFdRef, BufferSpec
-from tinygrad.helpers import getenv, to_mv, round_up, data64_le, all_same, flatten, DEBUG, AMD_LLVM, PROFILE, ProfileEvent, suppress_finalizing
-from tinygrad.helpers import lo32, hi32
+from tinygrad.helpers import getenv, to_mv, round_up, data64_le, DEBUG, AMD_LLVM, PROFILE, ProfileEvent, suppress_finalizing, lo32, hi32
 from tinygrad.renderer.cstyle import AMDRenderer
 from tinygrad.renderer.llvmir import AMDLLVMRenderer
 from tinygrad.runtime.autogen import kfd, hsa, pci, sqtt
@@ -109,17 +108,6 @@ class AMDComputeQueue(HWQueue):
 
     self.pkt3(self.pm4.PACKET3_RELEASE_MEM, event_dw | cache_flags_dw, memsel_dw, *data64_le(address), *data64_le(value), ctxid)
 
-  def xcc_barrier(self):
-    if self.dev.xcc_sync is None: return self
-    assert self.dev.xccs == 8, 'only 8 XCCs supported'
-    a, b = self.dev.xcc_sync
-    mem_eq = self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ) | self.pm4.WAIT_REG_MEM_MEM_SPACE(1)
-    self.pkt3(self.pm4.PACKET3_ATOMIC_MEM, self.soc.TC_OP_ATOMIC_ADD_RTN_32, *data64_le(a.value_addr), *data64_le(1), *data64_le(0), 0x10) # a += 1
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, mem_eq, *data64_le(a.value_addr), 0, 0b111, 0x80) # a == 0 (mod 8) via bitmask
-    self.pkt3(self.pm4.PACKET3_ATOMIC_MEM, self.soc.TC_OP_ATOMIC_ADD_RTN_32, *data64_le(b.value_addr), *data64_le(1), *data64_le(0), 0x10) # b += 1
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, mem_eq, *data64_le(b.value_addr), 0, 0b111, 0x80) # b == 0 (mod 8) via bitmask
-    return self
-
   def memory_barrier(self):
     pf = '' if self.nbio.version[0] == 2 else '0' if self.nbio.version[:2] != (7, 11) else '1'
     self.wait_reg_mem(reg_req=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
@@ -127,13 +115,6 @@ class AMDComputeQueue(HWQueue):
     self.acquire_mem()
     return self
 
-  def xcc_config(self):
-    self.wreg(self.gc.regCOMPUTE_TG_CHUNK_SIZE, 1)
-    for xcc_id in range(self.dev.xccs):
-      with self.pred_exec(xcc_mask=1 << xcc_id):
-        self.wreg(self.dev.regCOMPUTE_CURRENT_LOGIC_XCC_ID, xcc_id)
-    return self
-
   def spi_config(self, tracing:bool):
     self.wreg(self.gc.regSPI_CONFIG_CNTL, ps_pkr_priority_cntl=3, exp_priority_order=3, gpr_write_priority=0x2c688,
               enable_sqg_bop_events=int(tracing), enable_sqg_top_events=int(tracing))
@@ -278,16 +259,10 @@ class AMDComputeQueue(HWQueue):
 
     if prg.dev.sqtt_enabled: self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_MARKER) | self.pm4.EVENT_INDEX(0))
     self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.CS_PARTIAL_FLUSH) | self.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))
-
-    if self.dev.xccs > 1:
-      self.release_mem(cache_flush=True)
-      self.acquire_mem(gli=0)
-      self.xcc_barrier()
     return self
 
   def wait(self, signal:AMDSignal, value:sint=0):
     self.wait_reg_mem(mem=signal.value_addr, value=value, mask=0xffffffff)
-    if self.dev.xccs > 1 and not self.dev.is_aql: self.xcc_barrier()
     return self
 
   def timestamp(self, signal:AMDSignal):
@@ -464,14 +439,14 @@ class AMDProgram(HCQProgram):
     # TODO; this API needs the type signature of the function and global_size/local_size
     self.dev, self.name, self.lib = dev, name, lib
 
-    image, sections, _ = elf_loader(self.lib)
+    image, sections, relocs = elf_loader(self.lib)
 
     rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"), -1)
-    text_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".text"), -1)
-    assert rodata_entry >= 0 and text_entry >= 0, ".text or .rodata section not found"
+    assert rodata_entry >= 0, ".rodata section not found"
 
-    # Relo for kernel_code_entry_byte_offset for AMD_LLVM. Comgr doesn't need that, but keep shared code path.
-    image[rodata_entry+0x10:rodata_entry+0x10+8] = struct.pack('<q', text_entry - rodata_entry)
+    for apply_image_offset, rel_sym_offset, typ, addent in relocs:
+      if typ == 5: image[apply_image_offset:apply_image_offset+8] = struct.pack('<q', rel_sym_offset - apply_image_offset + addent) # R_AMDGPU_REL64
+      else: raise RuntimeError(f"unknown AMD reloc {typ}")
 
     self.lib_gpu = self.dev.allocator.alloc(round_up(image.nbytes, 0x1000), buf_spec:=BufferSpec(cpu_access=True, nolru=True))
     self.dev.allocator._copyin(self.lib_gpu, image)
@@ -538,12 +513,6 @@ class AMDQueueDesc:
   @property
   def read_ptr(self): return min(p[0] for p in self.read_ptrs)
 
-  @classmethod
-  def multi(cls, *queues: AMDQueueDesc):
-    assert all_same([(q.ring.addr, q.put_value) for q in queues]), f"All queues must have the same ring and put_value: {queues}"
-    return cls(ring=queues[0].ring, put_value=queues[0].put_value, doorbells=flatten(q.doorbells for q in queues),
-               read_ptrs=flatten(q.read_ptrs for q in queues), write_ptrs=flatten(q.write_ptrs for q in queues))
-
   def signal_doorbell(self, dev, doorbell_value:int|None=None):
     for write_ptr in self.write_ptrs: write_ptr[0] = self.put_value
 
@@ -707,14 +676,14 @@ class PCIIface(PCIIfaceBase):
 
   def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
     assert cwsr_buffer is None, "no cwsr buffer for am"
-    assert queue_type != kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL, "no AQL queues for am"
 
     if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA:
       self.dev_impl.sdma.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
                                     doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_sDMA_ENGINE0), pipe=0, queue=0)
     else:
       self.dev_impl.gfx.setup_ring(ring_addr=ring.va_addr, ring_size=ring.size, rptr_addr=gart.va_addr+rptr, wptr_addr=gart.va_addr+wptr,
-        eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0)
+        eop_addr=eop_buffer.va_addr, eop_size=eop_buffer.size, doorbell=(doorbell_index:=am.AMDGPU_NAVI10_DOORBELL_MEC_RING0), pipe=0, queue=0,
+        aql=(queue_type==kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL))
 
     return AMDQueueDesc(ring=ring.cpu_view().view(fmt='I'), doorbells=[self.dev_impl.doorbell64.view(doorbell_index * 8, 8, fmt='Q')],
       read_ptrs=[gart.cpu_view().view(offset=rptr, size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=wptr, size=8, fmt='Q')])
@@ -800,14 +769,11 @@ class AMDDevice(HCQCompiled):
     self.sdma = import_module('sdma', min(self.iface.ip_versions[am.SDMA0_HWIP], (6, 0, 0)))
     self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP], self.iface.ip_offsets[am.GC_HWIP])
 
-    # Define the regCOMPUTE_CURRENT_LOGIC_XCC_ID register, which is missing from the asic_regs files.
-    if self.target[:2] in {(9,4),(9,5)}: self.regCOMPUTE_CURRENT_LOGIC_XCC_ID = AMDReg("regCOMPUTE_CURRENT_LOGIC_XCC_ID", 0xe25, 0, {}, self.gc.bases)
-
     nbio_name = 'nbio' if self.target[0] < 12 else 'nbif'
     nbio_pad = (0,) if self.target[0] == 9 else ()
     self.nbio = AMDIP(nbio_name, self.iface.ip_versions[am.NBIF_HWIP], {i:nbio_pad+x for i,x in self.iface.ip_offsets[am.NBIF_HWIP].items()})
 
-    self.is_aql = getenv("AMD_AQL", 0)
+    self.is_aql = getenv("AMD_AQL", int(self.xccs > 1))
     if self.is_aql:
       self.pm4_ibs = self.iface.alloc(0x2000 if self.is_usb() else (16 << 20), uncached=True, cpu_access=True)
       self.pm4_ib_alloc = BumpAllocator(self.pm4_ibs.size, wrap=True)
@@ -829,13 +795,6 @@ class AMDDevice(HCQCompiled):
     self.max_private_segment_size = 0
     self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
 
-    # XCC setup
-    self.xcc_sync: tuple[AMDSignal, AMDSignal]|None = None
-    if self.xccs > 1 and not self.is_aql:
-      self.xcc_sync_area = self.allocator.alloc(0x1000, BufferSpec(nolru=True, cpu_access=True))
-      self.xcc_sync = (AMDSignal(base_buf=self.xcc_sync_area), AMDSignal(base_buf=self.xcc_sync_area.offset(256)))
-      cast(AMDComputeQueue, self.hw_compute_queue_t()).xcc_config().submit(self)
-
     # SQTT is disabled by default because of runtime overhead and big file sizes (~200mb to Tensor.full() two 4096x4096 tensors and matmul them)
     self.sqtt_enabled = PROFILE and bool(getenv("SQTT", 0))
     if self.sqtt_enabled:
@@ -866,10 +825,9 @@ class AMDDevice(HCQCompiled):
     cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
     eop_buffer = self.iface.alloc(eop_buffer_size) if eop_buffer_size else None
 
-    return AMDQueueDesc.multi(*(self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
-                                wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
-                                xcc_id=xcc_id, ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size)
-                                for xcc_id in range(self.xccs if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE else 1)))
+    return (self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+            wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
+            ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size))
 
   def _ensure_has_local_memory(self, required):
     if self.max_private_segment_size >= required: return

tinygrad/runtime/ops_null.py

@@ -1,28 +1,33 @@
+import functools
 from tinygrad.device import Compiled, Compiler, Allocator
 from tinygrad.engine.jit import MultiGraphRunner
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.uop.ops import Ops
+from tinygrad.helpers import cpu_profile
 
 class NullRenderer(CStyleLanguage):
   device = "NULL"
   has_local = False
   float4 = "float4"
+  barrier = "// BARRIER"
   code_for_op = {**CStyleLanguage.code_for_op, Ops.THREEFRY: lambda a,b,dtype: f"threefry({a},{b})", Ops.MAX: lambda a,b,dtype: f"max({a},{b})"}
 
 class NullProgram:
-  def __init__(self, name:str, lib:bytes): pass
+  def __init__(self, device:str, name:str, lib:bytes): self.device, self.name = device, name
   def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int, ...]=(), wait=False):
-    return 1e-4
+    with cpu_profile(self.name, self.device): return 1e-4
 
 class NullAllocator(Allocator['NullDevice']):
   def _alloc(self, size, options): pass
   def _copyin(self, dest, src:memoryview): pass
   def _copyout(self, dest:memoryview, src): pass
-  def _transfer(self, dest, src, sz:int, src_dev, dest_dev): pass
+  def _transfer(self, dest, src, sz:int, src_dev, dest_dev):
+    with cpu_profile(f"{src_dev.device} -> {dest_dev.device}", self.dev.device): pass
   def _offset(self, buf, offset:int, size:int): pass
 
 class NullGraph(MultiGraphRunner):
   def __call__(self, input_rawbuffers, var_vals, wait=False) -> float|None: return 1e-3
 
 class NullDevice(Compiled):
-  def __init__(self, device:str): super().__init__(device, NullAllocator(self), NullRenderer(), Compiler(), NullProgram, NullGraph)
+  def __init__(self, device:str): super().__init__(device, NullAllocator(self), NullRenderer(), Compiler(), functools.partial(NullProgram, device),
+                                                   NullGraph)

tinygrad/runtime/ops_python.py

@@ -4,25 +4,35 @@
 # this is the (living) definition of uops
 from typing import Any, TYPE_CHECKING
 import pickle, base64, itertools, time, struct, sys
-from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate
+from tinygrad.dtype import DType, dtypes, ImageDType, PtrDType, truncate, float_to_bf16
 from tinygrad.helpers import all_same, getenv, flatten, get_single_element
 from tinygrad.device import Compiled, Compiler, Allocator
 from tinygrad.codegen.opt import tc
 from tinygrad.uop.ops import exec_alu, Ops, UOp, GroupOp
 from tinygrad.renderer import Renderer
 
-def _load(m, i):
+def storage_fmt_for_dtype(dtype: DType): return 'H' if dtype == dtypes.bfloat16 else dtype.fmt
+
+def to_storage_scalar(x, dtype: DType):
+  if dtype == dtypes.bfloat16: return (struct.unpack('I', struct.pack('f', float_to_bf16(x)))[0] >> 16) & 0xFFFF
+  return x
+
+def from_storage_scalar(x, dtype: DType):
+  if dtype == dtypes.bfloat16: return struct.unpack('f', struct.pack('I', (x & 0xFFFF) << 16))[0]
+  return x
+
+def _load(m, i, dtype: DType):
   if i is None: return 0.0
   if i < 0 or i >= len(m): raise IndexError(f"load out of bounds, size is {len(m)} and access is {i}")
-  return m[i]
+  return from_storage_scalar(m[i], dtype)
 
-def load(inp, j=0):
-  if len(inp) == 2: return [_load(m, x+j if x is not None else None) if gate else default for (m,x,gate),default in zip(*inp)]
-  return [_load(m, x+j if x is not None else None) for m,x,_ in inp[0]]
+def load(inp, j, dtype: DType):
+  if len(inp) == 2: return [_load(m, x+j if x is not None else None, dtype) if gate else default for (m,x,gate),default in zip(*inp)]
+  return [_load(m, x+j if x is not None else None, dtype) for m,x,_ in inp[0]]
 
-def _store(m, i, v):
+def _store(m, i, v, dtype: DType):
   if i < 0 or i >= len(m): raise IndexError(f"store out of bounds, size is {len(m)}, access is {i}, value is {v}")
-  m[i] = v
+  m[i] = to_storage_scalar(v, dtype)
 
 class PythonProgram:
   def __init__(self, name:str, lib:bytes):
@@ -57,19 +67,20 @@ class PythonProgram:
         if uop is Ops.STORE:
           for j,val in enumerate(inp[1] if dtp[1].count > 1 else [inp[1]]):
             for (m,o,g),v in zip(inp[0], val):
-              if g: _store(m, o+j, v)
+              if g: _store(m, o+j, v, dtp[1].scalar())
           i += 1
           continue
         if uop in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG}:
           assert isinstance(dtype, PtrDType), dtype
-          if dtype.fmt is None: raise RuntimeError(f"{dtype=} is not supported")
-          if TYPE_CHECKING or sys.version_info < (3, 12): assert dtype.fmt != "e"
+          storage_fmt = storage_fmt_for_dtype(dtype.base.scalar())
+          if storage_fmt is None: raise RuntimeError(f"{dtype=} is not supported")
+          if TYPE_CHECKING or sys.version_info < (3, 12): assert storage_fmt != "e"
           if uop is Ops.DEFINE_REG:
             # REGs are per thread
-            ul[i] = [memoryview(bytearray(dtype.size*dtype.itemsize)).cast(dtype.fmt) for _ in range(warp_size)]
+            ul[i] = [memoryview(bytearray(dtype.size*dtype.itemsize)).cast(storage_fmt) for _ in range(warp_size)]
           else:
             buf = memoryview(bytearray(dtype.size*dtype.itemsize)) if uop is not Ops.DEFINE_GLOBAL else pbufs.pop(0)
-            ul[i] = [buf.cast(dtype.fmt)] * warp_size
+            ul[i] = [buf.cast(storage_fmt)] * warp_size
         elif uop is Ops.DEFINE_VAR:
           ul[i] = [pvals.pop(0)] * warp_size
         elif uop is Ops.SPECIAL:
@@ -98,16 +109,17 @@ class PythonProgram:
               continue
         elif uop is Ops.VECTORIZE: ul[i] = inp
         elif uop is Ops.BITCAST:
-          assert dtp[0].fmt and dtype.fmt
-          pack_format, unpack_format = str(warp_size) + dtp[0].fmt, str(warp_size) + dtype.fmt
-          ul[i] = list(struct.unpack(unpack_format, struct.pack(pack_format, *inp[0])))
+          packed = struct.pack(str(warp_size) + storage_fmt_for_dtype(dtp[0].scalar()), *[to_storage_scalar(x, dtp[0].scalar()) for x in inp[0]])
+          ul[i] = list(struct.unpack(str(warp_size) +  storage_fmt_for_dtype(dtype.scalar()), packed))
+          ul[i] = [from_storage_scalar(x, dtype.scalar()) for x in ul[i]]
         elif uop is Ops.CAST:
           ul[i] = [truncate.get(dtype, lambda dt: dt)(dtypes.as_const(x, dtype)) for x in inp[0]]
         elif uop is Ops.LOAD:
           if dtype.count > 1:
-            ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j) for j in range(dtype.count)]
+            ul[i] = [load([inp[i][j] if i != 0 and dtp[i].count > 1 else inp[i] for i in range(len(inp))], j, dtype.scalar()) \
+              for j in range(dtype.count)]
           else:
-            ul[i] = load(inp)
+            ul[i] = load(inp, 0, dtype)
         elif uop is Ops.GEP: ul[i] = inp[0][get_single_element(arg)]
         elif uop is Ops.WMMA:
           # here are the models for the WMMA instruction on the different hardware

tinygrad/runtime/support/am/ip.py

@@ -224,7 +224,8 @@ class AM_GFX(AM_IP):
     self._grbm_select()
     self.adev.regGCVM_CONTEXT0_CNTL.write(0)
 
-  def setup_ring(self, ring_addr:int, ring_size:int, rptr_addr:int, wptr_addr:int, eop_addr:int, eop_size:int, doorbell:int, pipe:int, queue:int):
+  def setup_ring(self, ring_addr:int, ring_size:int, rptr_addr:int, wptr_addr:int, eop_addr:int, eop_size:int, doorbell:int, pipe:int, queue:int,
+                 aql:bool):
     mqd = self.adev.mm.valloc(0x1000, uncached=True, contiguous=True)
 
     struct_t = getattr(am, f"struct_v{self.adev.ip_ver[am.GC_HWIP][0]}_compute_mqd")
@@ -235,9 +236,10 @@ class AM_GFX(AM_IP):
       cp_hqd_pq_rptr_report_addr_lo=lo32(rptr_addr), cp_hqd_pq_rptr_report_addr_hi=hi32(rptr_addr),
       cp_hqd_pq_wptr_poll_addr_lo=lo32(wptr_addr), cp_hqd_pq_wptr_poll_addr_hi=hi32(wptr_addr),
       cp_hqd_pq_doorbell_control=self.adev.regCP_HQD_PQ_DOORBELL_CONTROL.encode(doorbell_offset=doorbell*2, doorbell_en=1),
-      cp_hqd_pq_control=self.adev.regCP_HQD_PQ_CONTROL.encode(rptr_block_size=5, unord_dispatch=0, queue_size=(ring_size//4).bit_length()-2),
+      cp_hqd_pq_control=self.adev.regCP_HQD_PQ_CONTROL.encode(rptr_block_size=5, unord_dispatch=0, queue_size=(ring_size//4).bit_length()-2,
+        **({'queue_full_en':1, 'slot_based_wptr':2, 'no_update_rptr':1} if aql else {})),
       cp_hqd_ib_control=self.adev.regCP_HQD_IB_CONTROL.encode(min_ib_avail_size=0x3), cp_hqd_hq_status0=0x20004000,
-      cp_mqd_control=self.adev.regCP_MQD_CONTROL.encode(priv_state=1), cp_hqd_vmid=0,
+      cp_mqd_control=self.adev.regCP_MQD_CONTROL.encode(priv_state=1), cp_hqd_vmid=0, cp_hqd_aql_control=int(aql),
       cp_hqd_eop_base_addr_lo=lo32(eop_addr>>8), cp_hqd_eop_base_addr_hi=hi32(eop_addr>>8),
       cp_hqd_eop_control=self.adev.regCP_HQD_EOP_CONTROL.encode(eop_size=(eop_size//4).bit_length()-2))
 

tinygrad/runtime/support/memory.py

@@ -77,11 +77,10 @@ class TLSFAllocator:
       if self.lv1_entries[l1] == 0: continue
       for l2 in range(self.lv2(size) if l1 == size.bit_length() else 0, (1 << self.l2_cnt)):
         if len(self.storage[l1][l2]) > 0:
-          nsize = self.blocks[self.storage[l1][l2][0]][0]
-          assert nsize >= size, "block must be larger"
-
           # Block start address.
           start = self.storage[l1][l2][0]
+          nsize = self.blocks[start][0]
+          assert nsize >= size, "block must be larger"
 
           # If request contains alignment, split the block into two parts.
           if (new_start:=round_up(start, align)) != start:

tinygrad/runtime/support/nv/nvdev.py

@@ -118,7 +118,7 @@ class NVDev(PCIDevImplBase):
 
     self.include("src/common/inc/swref/published/turing/tu102/dev_fb.h")
     if self.reg("NV_PFB_PRI_MMU_WPR2_ADDR_HI").read() != 0:
-      if DEBUG >= 2: print(f"nv {self.devfmt}: WPR2 is up. Issuing a full reset.")
+      if DEBUG >= 2: print(f"nv {self.devfmt}: WPR2 is up. Issuing a full reset.", flush=True)
       System.pci_reset(self.devfmt)
       time.sleep(0.5)
 

tinygrad/schedule/kernelize.py

@@ -119,7 +119,7 @@ def create_kernel(x:UOp, b:UOp|None=None):
   if b is None: b = UOp.new_buffer(x.device, x.size, x.dtype)
   kernel = UOp(Ops.KERNEL, src=(b,)+x.src, arg=Kernel(x.sink(), m if (m:=x.metadata) else ()))
   buffer = b.base if b.size == b.base.size else UOp(Ops.BUFFER_VIEW, b.dtype, (b.base,), (b.size, b.arg.views[0].offset))
-  return buffer.assign(kernel).reshape(x.shape)
+  return buffer.assign(kernel).shrink(((0, prod(x.shape)),)).reshape(x.shape)
 
 DONT_PLACE_IN_KERNEL = {Ops.KERNEL, Ops.ASSIGN, Ops.BUFFER, Ops.MSELECT, Ops.MSTACK, Ops.MULTI, Ops.BIND}
 def append_to_kernel(x:UOp):

tinygrad/schedule/rangeify.py

@@ -1,12 +1,12 @@
 from typing import Any
 from dataclasses import dataclass, field
-from tinygrad.dtype import dtypes, PtrDType
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady, _substitute, AxisType
-from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, colored, RANGEIFY
+from tinygrad.dtype import dtypes, PtrDType, ImageDType, AddrSpace
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, RewriteNotReady, _substitute
+from tinygrad.helpers import argsort, prod, all_same, pluralize, getenv, RANGEIFY
 from tinygrad.schedule.multi import multi_pm
 
 from tinygrad.schedule.kernelize import Kernel
-from tinygrad.uop.ops import track_rewrites, graph_rewrite_map, graph_rewrite, KernelInfo, identity_element, sint
+from tinygrad.uop.ops import track_rewrites, graph_rewrite_map, graph_rewrite, identity_element, sint, AxisType
 
 # 0. do some cleanup rewrites, mostly copied from the old stuff
 
@@ -109,7 +109,7 @@ class RangeifyContext:
   # create ranges
   range_idx: int = 0
   def new_range(self, s:sint, axistype:AxisType=AxisType.LOOP):
-    ret = UOp.range(dtypes.int, s, self.range_idx, axistype)
+    ret = UOp.range(s, self.range_idx, axistype)
     self.range_idx += 1
     return ret
 
@@ -196,7 +196,8 @@ def map_contiguous(ctx:RangeifyContext, x:UOp):
   ranges = []
   for s in x.shape[len(x.src)-1:]:
     ranges.append(ctx.new_range(s) if resolve(s!=1) else UOp.const(dtypes.int, 0))
-  return x.src[0].index(*ranges).bufferize(*x.src[1:], *[x for x in ranges if x.op is not Ops.CONST], arg=x.device).forced_reshape(x.shape)
+  ret = x.src[0].index(*ranges).bufferize(*x.src[1:], *[x for x in ranges if x.op is not Ops.CONST], arg=x.device)
+  return ret.shrink(((0, prod(x.shape)),)).forced_reshape(x.shape)
 
 def map_reduce(ctx:RangeifyContext, idx:UOp, red:UOp):
   rngs = list(idx.src[1:])
@@ -329,18 +330,28 @@ pm_cleanups = double_reshape+pm_mops+PatternMatcher([
 # BUFFERIZE returns the BUFFER ready for INDEXing (doing this will make splitting a lot easier)
 # NOTE: this has been fixed up a bit
 
-def bufferize_to_store(x:UOp):
+def bufferize_to_store(x:UOp, locals_allowed=False):
   rngs = x.src[1:]
   shape = tuple([int(r.vmax+1) for r in rngs])
-  sdtype = x.dtype.ptr(size=prod(shape))
-  assert prod(shape) > 0, f"no zero sized buffers {shape}"
+  size = prod(shape)
+  assert size > 0, f"no zero sized buffers {shape}"
+  sdtype = x.dtype.ptr(size=size, addrspace=AddrSpace.GLOBAL if not isinstance(x.arg, tuple) else x.arg[0])
   if x.src[0].op is Ops.ASSIGN:
     assign_target, assign_src = x.src[0].src
     assert assign_target.op is Ops.INDEX
     return assign_target.replace(dtype=sdtype).store(assign_src, *rngs, dtype=sdtype)
-  buf = UOp.new_buffer(x.arg, prod(shape), x.dtype)
+  # NOTE: the DEFINE_LOCAL needs to be disambiguated here
+  if sdtype.addrspace == AddrSpace.GLOBAL:
+    buf = UOp.new_buffer(x.arg, size, x.dtype)
+  else:
+    if not locals_allowed: return None
+    buf = UOp(Ops.DEFINE_LOCAL, sdtype, arg=x.arg[1])
   return buf.reshape(shape).index(*rngs, dtype=sdtype).store(x.src[0], *rngs, dtype=sdtype).forced_reshape(shape, dtype=x.dtype)
 
+pm_add_buffers_local = pm_mops+PatternMatcher([
+  (UPat(Ops.BUFFERIZE, name="x"), lambda x: bufferize_to_store(x, True)),
+])
+
 pm_add_buffers = pm_mops+PatternMatcher([
   (UPat(Ops.BUFFERIZE, name="x"), bufferize_to_store),
 
@@ -380,31 +391,33 @@ to_define_global = PatternMatcher([
   (UPat(Ops.BIND, name="b"), unbind_kernel),
   (UPat((Ops.ASSIGN, Ops.MSTACK, Ops.MSELECT), name="assign"), handle_assign),
 
-  # add loads to non ptr indexes
-  # TODO: this can be moved into codegen?
-  (UPat((Ops.DEFINE_GLOBAL, Ops.STORE), name="dg").f(Ops.INDEX, name="idx", allow_any_len=True),
-   lambda dg,idx: idx.replace(dtype=dg.dtype, arg=None).load() if not isinstance(idx.dtype, PtrDType) else None),
-
-  # TODO: this can be moved into codegen
-  (UPat(Ops.STORE, name="store").f(Ops.INDEX, allow_any_len=True, name="idx").f(Ops.LOAD),
-    lambda store,idx: idx.replace(src=(store.as_buf(),)+idx.src[1:]).load(store)),
-
   # HACK in case any CONSTs were replaced
   # this is only needed if you are using symbolic
   #(UPat(Ops.CONST, name="c"), lambda c: c.replace(src=()) if len(c.src) else None),
 ])
 
+rangeify_codegen = PatternMatcher([
+  # add loads to non ptr indexes
+  # TODO: this can be moved into codegen?
+  (UPat((Ops.DEFINE_GLOBAL, Ops.STORE), name="dg").f(Ops.INDEX, name="idx", allow_any_len=True),
+   lambda dg,idx: None if isinstance(idx.dtype, (PtrDType, ImageDType)) else idx.replace(dtype=dg.dtype, arg=None).load()),
+
+  # TODO: this can be moved into codegen
+  (UPat(Ops.STORE, name="store").f(Ops.INDEX, allow_any_len=True, name="idx").f(Ops.LOAD),
+    lambda store,idx: idx.replace(src=(store.as_buf(),)+idx.src[1:]).load(store if idx.dtype.addrspace != AddrSpace.LOCAL else store.barrier())),
+
+  # TODO: hack for group for reduce
+  (UPat(Ops.IF, src=(UPat.var("gate"), UPat(Ops.LOAD, src=(UPat.var("src"), UPat.var("barrier"))),)),
+   lambda src, barrier, gate: src.load(UOp(Ops.IF, src=(gate, barrier)))),
+])
+
 def split_store(x:UOp):
   if len(x.ranges): return None
   ctx = LocalAddBufferContext()
-  ret = graph_rewrite(x, to_define_global, ctx=ctx, name="kernel split", bottom_up=True)
-
-  store_rngs = ret.src[2:]
-  rng = sorted([u for u in ret.toposort() if u.op is Ops.RANGE], key=lambda x: x.arg)
-  name = "k"+colored('_', 'BLACK').join(['']+[colored(s.src[0].render(), "WHITE" if s in store_rngs else "red") for s in rng])
+  ret = graph_rewrite(x, to_define_global+rangeify_codegen, ctx=ctx, name="kernel split", bottom_up=True)
 
   # NOTE: the hack for COPY is here
-  ret = ret.sink(arg=KernelInfo(name=name)) if ret.src[1].op is not Ops.COPY else ret.src[1]
+  ret = ret.sink() if ret.src[1].op is not Ops.COPY else ret.src[1]
   kernel = UOp(Ops.KERNEL, src=tuple(ctx.map.values())+tuple(ctx.vars.keys()), arg=Kernel(ret,()))
   return x.as_buf().assign(kernel)
 

tinygrad/shape/view.py

@@ -3,7 +3,7 @@ import functools, operator, itertools
 from dataclasses import dataclass
 from typing import cast, Sequence
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import resolve, UOp, Variable, sint, sym_infer, smax, smin, sint_to_uop, Ops, ssimplify
+from tinygrad.uop.ops import resolve, UOp, Variable, sint, smax, smin, sint_to_uop, Ops, ssimplify
 from tinygrad.helpers import prod, all_int, argsort, flatten, ceildiv
 
 # returns the axes to create new_shape if new_shape can be created by combining axis from old_shape
@@ -114,7 +114,7 @@ class View:
 
   def to_indexed_uops(self:View, idxs:Sequence[UOp]|None=None, vexpr:UOp=UOp.const(dtypes.bool, True)) -> tuple[UOp, UOp]:
     """(idx, valid)"""
-    if idxs is None: idxs = [UOp.range(dtypes.int, s, i) for i,s in enumerate(self.shape)]
+    if idxs is None: idxs = [UOp.range(s, i) for i,s in enumerate(self.shape)]
     iexpr = sint_to_uop(self.offset)
     for idx,sh,st,m in zip(idxs, self.shape, self.strides, self.mask if self.mask is not None else itertools.repeat(None)):
       if resolve(sh != 1) and resolve(st != 0): iexpr = iexpr + idx*st
@@ -311,9 +311,10 @@ class View:
 
     if not all(x >= 0 for x in new_shape): raise ValueError(f"shape can't contain negative numbers {new_shape}")
     # check for the same size
-    if (self_all_int := all_int(self.shape)):
-      assert all(isinstance(s, (int, UOp)) for s in new_shape), f"{self.shape=} -> {new_shape=} contains non (int, Variable) dim"
-      if resolve(prod(self.shape) != prod(new_shape), False): raise ValueError(f"size mismatched, can't reshape {self.shape=} -> {new_shape=}")
+    if all_int(self.shape):
+      # reshapes cannot introduce symbolic shape
+      assert all_int(new_shape), f"{self.shape=} -> {new_shape=} contains non int dims"
+      if prod(self.shape) != prod(new_shape): raise ValueError(f"size mismatched, can't reshape {self.shape=} -> {new_shape=}")
 
     if 0 in self.shape: return View.create(new_shape)
     if new_shape == () and self.mask and any(mx==my for (mx,my) in self.mask): return None
@@ -321,15 +322,6 @@ class View:
     # after the asserts, it's okay to check contiguous
     if self.contiguous: return View.create(new_shape)
 
-    # if it's not contiguous and new shape is symbolic, check if it's directly replaceable
-    if self_all_int and not all_int(new_shape):
-      if len(self.shape) != len(new_shape): raise ValueError(f"cannot symbolic reshape non-contiguous {self} -> {new_shape}")
-      for si, so in zip(self.shape, new_shape):
-        if not isinstance(so, int): so = sym_infer(so, dict([v.unbind() for v in so.vars()]))
-        if si != so: raise ValueError(f"cannot symbolic reshape non-contiguous {self} -> {new_shape}")
-      # all dimensions matched, return the new view directly
-      return View(new_shape, self.strides, self.offset, self.mask, self.contiguous)
-
     r_strides, r_new_shape = [], reversed(new_shape)
     for merged_size, new_stride, real_size in reversed(merge_dims(self.shape, self.strides, self.mask)):
       # TODO: write with get_contraction

tinygrad/tensor.py

@@ -6,7 +6,7 @@ from typing import Callable, ClassVar, Sequence, cast, get_args, Literal, Suppor
 from tinygrad.dtype import DType, DTypeLike, dtypes, ImageDType, ConstType, least_upper_float, least_upper_dtype, sum_acc_dtype, to_dtype, truncate
 from tinygrad.dtype import _from_np_dtype, _to_np_dtype
 from tinygrad.helpers import argfix, make_tuple, flatten, prod, all_int, round_up, merge_dicts, argsort, getenv, all_same, fully_flatten, dedup
-from tinygrad.helpers import IMAGE, WINO, Metadata, TRACEMETA, ceildiv, fetch, polyN, unwrap, DEBUG, is_numpy_ndarray, RANGEIFY
+from tinygrad.helpers import IMAGE, WINO, Metadata, TRACEMETA, ceildiv, fetch, polyN, unwrap, DEBUG, is_numpy_ndarray, RANGEIFY, FUSE_ATTENTION
 from tinygrad.gradient import compute_gradient
 from tinygrad.uop.ops import smax, smin, resolve, UOp, Ops, sint, Variable, MathTrait, identity_element, all_metadata
 from tinygrad.uop.spec import tensor_uop_spec, type_verify
@@ -68,7 +68,7 @@ def _frompy(x:list|tuple|bytes, dtype:DType) -> UOp:
     ret = UOp.new_buffer("PYTHON", prod(shape:=get_shape(x)), dtype).reshape(shape)
     assert dtype.fmt is not None, f"{dtype=} has None fmt"
     truncate_function = truncate[dtype]
-    data = struct.pack(f"@{ret.size}{dtype.fmt}", *[truncate_function(xi) for xi in fully_flatten(x)])
+    data = struct.pack(f"{ret.size}{dtype.fmt}", *[truncate_function(dtypes.as_const(xi, dtype)) for xi in fully_flatten(x)])
   # fake realize
   ret.buffer.allocate(memoryview(data if Device.DEFAULT != "PYTHON" else bytearray(data)))
   return ret
@@ -442,7 +442,7 @@ class Tensor(MathTrait):
     if not isinstance(size:=prod([x.vmax if isinstance(x, UOp) else x for x in shape]), int): raise ValueError(f"size must be int {size}")
     # TODO: add test for multidevice tensor
     device = tuple(Device.canonicalize(d) for d in device) if isinstance(device, tuple) else Device.canonicalize(device)
-    return Tensor(UOp.new_buffer(device, size, dtype), device, dtype, **kwargs).reshape(shape)
+    return Tensor(UOp.new_buffer(device, size, dtype), device, dtype, **kwargs).shrink(((0,prod(shape)),)).reshape(shape)
 
   @staticmethod
   def from_blob(ptr:int, shape:tuple[int, ...], **kwargs) -> Tensor:
@@ -3930,7 +3930,11 @@ class Tensor(MathTrait):
     if enable_gqa:
       key = key.repeat_interleave(self.shape[-3] // key.shape[-3], dim=-3)
       value = value.repeat_interleave(self.shape[-3] // value.shape[-3], dim=-3)
-    qk = self.matmul(key.transpose(-2,-1), dtype=least_upper_dtype(self.dtype, key.dtype, dtypes.float32)) / math.sqrt(self.shape[-1])
+
+    if FUSE_ATTENTION: q, key, value = self.contiguous(), key.contiguous(), value.contiguous()
+    else: q = self
+
+    qk = q.matmul(key.transpose(-2,-1), dtype=least_upper_dtype(q.dtype, key.dtype, dtypes.float32)) / math.sqrt(q.shape[-1])
     # handle attention mask
     if is_causal:
       if attn_mask is not None: raise RuntimeError("cannot set attn_mask when is_causal=True")
@@ -3938,7 +3942,8 @@ class Tensor(MathTrait):
     if attn_mask is not None:
       if attn_mask.dtype == dtypes.bool: attn_mask = attn_mask.where(0, -float("inf"))
       qk = qk + attn_mask
-    return qk.cast(self.dtype).softmax(-1).dropout(dropout_p) @ value
+    attn = qk.cast(self.dtype).softmax(-1).dropout(dropout_p) @ value
+    return attn.fuse() if FUSE_ATTENTION else attn
 
   def _do_reduction(self, reduction:ReductionStr="mean") -> Tensor:
     if reduction not in get_args(ReductionStr): raise ValueError(f"{reduction=} must be one of {get_args(ReductionStr)}")

tinygrad/uop/decompositions.py

@@ -280,7 +280,7 @@ def magicgu(vmax:int, d:int) -> tuple[int,int]:
       return m, s
   assert False
 
-def fast_idiv(device: str, x: UOp, d: int) -> UOp|None:
+def fast_idiv(device: str, x: UOp, d: int, dont_cast=False) -> UOp|None:
   # If d is a power of two this is not valid for signed ints!
   is_unsigned = True if x.vmin>=0 or x.dtype in dtypes.uints else False
   assert d>0, "Sign should have been taken out of divisor"
@@ -288,6 +288,10 @@ def fast_idiv(device: str, x: UOp, d: int) -> UOp|None:
   m,s = magicgu(max(vmax, abs(vmin)), d)
   if m*vmin >= dtypes.min(x.dtype) and m*vmax <= dtypes.max(x.dtype):
     return ((x*m) >> s) if is_unsigned else ((x*m) >> s) + (x<0).where(x.ufix(1), 0)
+  # before we try casting to a larger dtype (slow), we see if there are powers of two in d we can shift to make x smaller
+  if (largest_factor_of_two_in_d := (d & -d)) > 1:
+    if (ret:=fast_idiv(device, x//largest_factor_of_two_in_d, d//largest_factor_of_two_in_d, dont_cast=True)) is not None: return ret
+  if dont_cast: return None
   # promo_lattice needs to return an unsigned type if the type is unsigned
   if dtypes.is_int(next_dtype := promo_lattice[x.dtype][-1]) and is_dtype_supported(next_dtype, None if device=='' else device):
     if m*vmin >= dtypes.min(next_dtype) and m*vmax <= dtypes.max(next_dtype):
@@ -329,6 +333,8 @@ def get_late_rewrite_patterns(ops:tuple[Ops, ...], force_transcendental=False):
   if Ops.SQRT not in ops: pat.append((UPat(Ops.SQRT, src=UPat.var("d")), lambda d: xpow(d, d.const_like(0.5))))
   # rewrite MOD to AND (which should always be supported, but not for generic in tests): x % (2**y) -> x & (2**y-1)
   if Ops.AND in ops: pat += [(UPat.var("x", dtypes.ints)%UPat.cvar("c"), lambda x,c: x & (c.arg-1) if c.arg in powers_of_two else None)]
+  if Ops.OR in ops: pat += [(UPat.var("x", dtypes.bool).logical_not()&UPat.var("y", dtypes.bool).logical_not(),
+    lambda x,y: (x | y).logical_not())]
   # rewrite MUL/IDIV to SHL+SHR: x*(2**y) -> shl(x,y) and x//(2**y) -> shr(x,y)
   if Ops.SHL in ops: pat += [(UPat.var("x", dtypes.ints)*UPat.cvar("c"), lambda c,x: x << v if (v:=powers_of_two.get(c.arg, 0)) else None)]
   if Ops.SHR in ops:

tinygrad/uop/ops.py

@@ -142,6 +142,7 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     if self.op is Ops.INDEX and self.src[0].op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG,
                                                    Ops.BUFFER, Ops.BUFFERIZE, Ops.VECTORIZE, Ops.STORE}:
       return None
+    if self.op is Ops.BARRIER: return None
     if self.op in GroupOp.Block: return None
     from tinygrad.shape.shapetracker import ShapeTracker
     # VIEW and MovementOps define a new ShapeTracker from the arg
@@ -202,17 +203,13 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
   @functools.cached_property
   def ranges(self) -> dict[UOp, None]:
     if self.op is Ops.RANGE: return {self:None}
-    if self.op in {Ops.BUFFERIZE, Ops.REDUCE}:
-      ret = self.src[0].ranges.copy()
-      for s in self.src[1:]:
-        if s in ret: del ret[s]
-    elif self.op in {Ops.STORE}:
-      ret = self.src[0].ranges.copy()
-      ret.update(self.src[1].ranges)
-      for s in self.src[2:]:
+    range_start = {Ops.BUFFERIZE: 1, Ops.REDUCE: 1, Ops.STORE: 2, Ops.WMMA: 3}
+    ret: dict[UOp, None] = {}
+    if self.op in range_start.keys():
+      for s in self.src[:range_start[self.op]]: ret.update(s.ranges)
+      for s in UOp.sink(*self.src[range_start[self.op]:]).ranges:
         if s in ret: del ret[s]
     else:
-      ret = {}
       for s in self.src: ret.update(s.ranges)
     return ret
 
@@ -251,7 +248,8 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
     ret = self.arg[1] if self.op is Ops.REDUCE_AXIS else self.arg[7]
     assert isinstance(ret, tuple) and all(isinstance(x, int) for x in ret), f"axis_arg trying to return {ret}"
     return ret
-  def sink(self, *srcs:UOp|None, **kwargs): return UOp(Ops.SINK, dtypes.void, (self,)+tuple([x for x in srcs if x is not None]), **kwargs)
+  def sink(*srcs:UOp|None, **kwargs):  # pylint: disable=no-self-argument
+    return UOp(Ops.SINK, dtypes.void, tuple([x for x in srcs if x is not None]), **kwargs)
   def detach(self): return UOp(Ops.DETACH, self.dtype, (self,))
   def index(self, *srcs:UOp|None, **kwargs):
     return UOp(Ops.INDEX, kwargs.pop("dtype", self.dtype), (self,)+tuple([x for x in srcs if x is not None]), **kwargs)
@@ -299,8 +297,10 @@ class UOp(MathTrait, metaclass=UOpMetaClass):
       else: ret = ret.replace(src=(UOp(Ops.DEVICE, arg=device),))
     return ret
   @staticmethod
-  def range(dtype:DType, end:sint, idx:int, axistype:AxisType=AxisType.LOOP):
-    return UOp(Ops.RANGE, dtype=dtype, src=(sint_to_uop(end),), arg=(idx, axistype))
+  def range(end:sint, *arg):
+    if len(arg) == 0: raise RuntimeError("range needs an arg")
+    if len(arg) == 1: arg = arg+(AxisType.LOOP,)
+    return UOp(Ops.RANGE, dtype=dtypes.int, src=(sint_to_uop(end),), arg=arg)
   def r(self, op:Ops, axis:tuple[int, ...]):
     axis = tuple(sorted([x for x in axis if resolve(self.shape[x] != 1)]))
     if len(axis) == 0: return self
@@ -681,7 +681,8 @@ class UPat(MathTrait):
   def var(name:str|None=None, dtype:DType|tuple[DType, ...]|None=None): return UPat(dtype=dtype, name=name)
   @staticmethod
   @functools.cache
-  def cvar(name:str|None=None, dtype:DType|None=None, vec=True): return UPat((Ops.CONST,Ops.VCONST) if vec else Ops.CONST, dtype, name=name)
+  def cvar(name:str|None=None, dtype:DType|tuple[DType, ...]|None=None, vec=True):
+    return UPat((Ops.CONST,Ops.VCONST) if vec else Ops.CONST, dtype, name=name)
   @staticmethod
   def const(dtype:DType|tuple[DType, ...]|None, b:ConstType): return UPat(Ops.CONST, dtype=dtype, arg=b)
 

tinygrad/uop/spec.py

@@ -17,33 +17,39 @@ try:
     return s
 
   # ctx is (solver, load_number_dict)
+  # each uop gets rewritten to NOOP(arg=(solver, z3_object)), the arg has the solver first due to UOpMetaClass caching. z3 objects from different
+  # contexts can have the same hash but error on comparison
   z3_renderer = PatternMatcher([
     # Ops.SPECIAL can have symbolic arg but it wont be in the toposort beacuse its not a src, we need to add it manually
     (UPat(Ops.SPECIAL, src=(), name="x"), lambda x: UOp(Ops.SPECIAL, arg=x.arg[0], src=(x.ufix(x.arg[1]),))),
-    (UPat(Ops.SPECIAL, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=create_bounded(x.arg, 0, x.src[0].arg-1, ctx[0]))),
-    (UPat(Ops.DEFINE_VAR, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=create_bounded(x.arg[0], x.arg[1], x.arg[2], ctx[0]))),
-    (UPat(Ops.RANGE, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=create_bounded(f"ridx{x.arg}", 0, x.src[0].arg-1, ctx[0]))),
+    (UPat(Ops.SPECIAL, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg, 0, x.src[0].arg[1]-1, ctx[0])))),
+    (UPat(Ops.DEFINE_VAR, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(x.arg[0], x.arg[1], x.arg[2], ctx[0])))),
+    (UPat(Ops.RANGE, name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"ridx{x.arg}", 0, x.src[0].arg[1]-1, ctx[0])))),
     # float loads only become a variable when they get cast to int/bool
     (UPat(Ops.LOAD, dtypes.ints, name="x"),
-      lambda x,ctx: UOp(Ops.NOOP, arg=create_bounded(f"load{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0]))),
+      lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],create_bounded(f"load{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0])))),
     (UPat(Ops.CONST, dtype=dtypes.ints+(dtypes.bool,), name="x"),
-      lambda x,ctx: UOp(Ops.NOOP, arg=(z3.BoolVal if dtypes.is_bool(x.dtype) else z3.IntVal)(x.arg, ctx=ctx[0].ctx))),
+      lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0],(z3.BoolVal if dtypes.is_bool(x.dtype) else z3.IntVal)(x.arg, ctx=ctx[0].ctx)))),
     # z3 can cast from bool to int automatically
     (UPat(Ops.CAST, dtype=dtypes.ints, src=UPat(Ops.NOOP), name="x"), lambda x: x.src[0]),
-    (UPat(Ops.CAST, dtype=dtypes.bool, src=UPat(Ops.NOOP), name="x"), lambda x: UOp(Ops.NOOP, arg=(x.src[0].arg!=0))),
+    (UPat(Ops.CAST, dtype=dtypes.bool, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], x.src[0].arg[1]!=0))),
     # if the source of the cast is not a noop it means that it is a float and so we create a new variable
     (UPat(Ops.CAST, dtype=dtypes.ints, name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=create_bounded(f"cast{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0]))),
+      UOp(Ops.NOOP, arg=(ctx[0], create_bounded(f"cast{ctx[1].setdefault(x, len(ctx[1]))}", x.dtype.min, x.dtype.max, ctx[0])))),
     (UPat(Ops.CAST, dtype=dtypes.bool, name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=z3.Bool(f"cast{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx))),
+      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"cast{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
     (UPat(Ops.XOR, src=UPat(Ops.NOOP), name="x"),
-      lambda x: UOp(Ops.NOOP, arg=z3.BV2Int(z3_alu[x.op](*(z3.Int2BV(s.arg, x.dtype.itemsize*8) for s in x.src))))),
-    (UPat(GroupOp.ALU, src=UPat(Ops.NOOP), name="x"), lambda x: UOp(Ops.NOOP, arg=z3_alu[x.op](*(s.arg for s in x.src)))),
+      lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], z3.BV2Int(z3_alu[x.op](*(z3.Int2BV(s.arg[1], x.dtype.itemsize*8) for s in x.src)))))),
+    (UPat(GroupOp.ALU, src=UPat(Ops.NOOP), name="x"), lambda x,ctx: UOp(Ops.NOOP, arg=(ctx[0], z3_alu[x.op](*(s.arg[1] for s in x.src))))),
     # A comparison between floats introduces a new bool variable
     (UPat(GroupOp.Comparison, src=UPat(dtype=dtypes.floats), name="x"), lambda x,ctx:
-      UOp(Ops.NOOP, arg=z3.Bool(f"float_cmp{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx))),
+      UOp(Ops.NOOP, arg=(ctx[0], z3.Bool(f"float_cmp{ctx[1].setdefault(x, len(ctx[1]))}",ctx=ctx[0].ctx)))),
   ])
 
+  def uops_to_z3(solver, *uops: UOp) -> 'list[z3.ExprRef]':
+    with Context(TRACK_MATCH_STATS=0):  # cant pickle z3 objects
+      return [s.arg[1] for s in graph_rewrite(uops[0].sink(*uops[1:]), z3_renderer, ctx=(solver, {})).src]
+
   z3_imported = True
 except (ImportError, AttributeError): z3_imported = False
 
@@ -124,9 +130,8 @@ def validate_index(idx:UOp, gate:UOp=UOp.const(dtypes.bool, True)):
 
   if not z3_imported: raise ImportError("z3 is required for bounds checking, try IGNORE_OOB=0 or \"pip install z3-solver\"")
   solver = z3.Solver(ctx=z3.Context())
-  z3_sink = graph_rewrite(idx.src[1].sink(mask), z3_renderer, ctx=(solver, {}))
-  z3_idx = z3_sink.src[0].arg
-  solver.add(z3_sink.src[1].arg)
+  z3_idx, z3_mask = uops_to_z3(solver, idx.src[1], mask)
+  solver.add(z3_mask)
   if solver.check((z3_idx<0)|(sz<=z3_idx)) == z3.sat:
     print(f"idx={idx.src[1].render(simplify=False)}")
     print(f"mask & gate={mask.render(simplify=False)}")

tinygrad/uop/symbolic.py

@@ -1,5 +1,5 @@
 # all of symbolic lives here now
-from typing import Any, cast
+from typing import cast
 import math, operator, struct, functools
 from collections import defaultdict
 from tinygrad.uop.ops import Ops, PatternMatcher, UPat, UOp, GroupOp, exec_alu
@@ -19,7 +19,7 @@ def simplify_pow(x:UOp, c:UOp) -> UOp|None:
 def fold_bitcast(root:UOp, c:UOp) -> UOp|None:
   if (from_fmt:=c.dtype.scalar().fmt) is None or (to_fmt:=root.dtype.scalar().fmt) is None: return None
   if c.dtype.itemsize != root.dtype.itemsize: return None
-  def convert(v:Any): return struct.unpack(to_fmt, struct.pack(from_fmt, v))[0]
+  def convert(v:ConstType): return struct.unpack(to_fmt, struct.pack(from_fmt, v))[0]
   return root.const_like(convert(c.arg) if root.dtype.count == 1 else tuple(map(convert, c.arg)))
 
 symbolic_simple = PatternMatcher([
@@ -291,6 +291,9 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   # alu of two where with same conds can combine, only do if true branch or false branch is const
   (UPat(GroupOp.Binary, name="alu", src=(UPat.var("c").where(UPat.var("t"), UPat.var("f")), UPat.var("c").where(UPat.var("tt"), UPat.var("ff")))), \
    lambda alu,c,t,tt,f,ff: c.where(t.alu(alu.op, tt), f.alu(alu.op, ff)) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
+  # if its a plus we add the associative variation too
+  ((UPat.var("y")+UPat.var("c").where(UPat.var("t"), UPat.var("f"))) + UPat.var("c").where(UPat.var("tt"), UPat.var("ff")), \
+   lambda y,c,t,tt,f,ff: y+c.where(t+tt, f+ff) if t.op == tt.op == Ops.CONST or f.op == ff.op == Ops.CONST else None),
   # ALU/variable min==max -> CONST (slow!)
   (UPat(GroupOp.ALU|{Ops.DEFINE_VAR, Ops.SPECIAL, Ops.RANGE}, name="x"), lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
   # max folding

tinygrad/viz/index.html

@@ -75,9 +75,14 @@
   g.tag circle {
     fill: #FFD700;
     stroke: #B8860B;
+  }
+  g.port circle {
+    fill: #b3dcc2;
+  }
+  g.tag circle, #edge-labels circle {
     stroke-width: 0.8;
   }
-  g.tag text {
+  g.tag text, #edge-labels text {
     text-anchor: middle;
     font-size: 6px;
     fill: #08090e;
@@ -85,11 +90,30 @@
   .label :is(text, p) {
     font-weight: 350;
   }
+  rect.node {
+    stroke-width: 1.4;
+    stroke: #4a4b57;
+  }
+  rect.overlay {
+    fill: rgba(26, 27, 38, 0.5);
+  }
   .edgePath {
     stroke: #4a4b57;
     fill: none;
     stroke-width: 1.4px;
   }
+  .highlight rect, .edgePath.highlight, g.port circle {
+    stroke: #89C9A2;
+  }
+  #edge-labels g.port.highlight {
+    display: block
+  }
+  #edge-labels g.port {
+    display: none
+  }
+  #arrowhead {
+    fill: #4a4b57;
+  }
   .main-container {
     display: flex;
     width: 100%;
@@ -331,7 +355,7 @@
         </g>
         <defs>
           <marker id="arrowhead" viewBox="0 -5 10 10" refX="10" refY="0" markerWidth="6" markerHeight="6" orient="auto">
-            <path d="M0,-5L10,0L0,5" fill="#4a4b57"></path>
+            <path d="M0,-5L10,0L0,5" fill="context-stroke"></path>
           </marker>
         </defs>
       </svg>

tinygrad/viz/js/index.js

@@ -4,6 +4,15 @@ const displayGraph = (cls) => {
   for (const e of document.getElementsByClassName("view")) e.style.display = e.classList.contains(cls) ? "flex" : "none";
 }
 
+const darkenHex = (h, p = 0) =>
+  `#${(
+    c = parseInt(h.slice(1), 16),
+    f = 1 - p / 100,
+    ((c >> 16 & 255) * f | 0) << 16 |
+    ((c >>  8 & 255) * f | 0) <<  8 |
+    ((c       & 255) * f | 0)
+  ).toString(16).padStart(6, '0')}`;
+
 const ANSI_COLORS = ["#b3b3b3", "#ff6666", "#66b366", "#ffff66", "#6666ff", "#ff66ff", "#66ffff", "#ffffff"];
 const parseColors = (name, defaultColor="#ffffff") => Array.from(name.matchAll(/(?:\u001b\[(\d+)m([\s\S]*?)\u001b\[0m)|([^\u001b]+)/g),
   ([_, code, colored_st, st]) => ({ st: colored_st ?? st, color: code != null ? ANSI_COLORS[(parseInt(code)-30+60)%60] : defaultColor }));
@@ -56,11 +65,23 @@ async function renderDag(graph, additions, recenter=false) {
     const g = dagre.graphlib.json.read(e.data);
     // draw nodes
     const STROKE_WIDTH = 1.4;
+    d3.select("#graph-svg").on("click", () => d3.selectAll(".highlight").classed("highlight", false));
     const nodes = d3.select("#nodes").selectAll("g").data(g.nodes().map(id => g.node(id)), d => d).join("g")
-      .attr("transform", d => `translate(${d.x},${d.y})`).classed("clickable", d => d.ref != null)
-      .on("click", (_,d) => setCtxWithHistory(d.ref));
+      .attr("transform", d => `translate(${d.x},${d.y})`).classed("clickable", d => d.ref != null).on("click", (e,d) => {
+        if (d.ref != null) return setCtxWithHistory(d.ref);
+        const parents = g.predecessors(d.id);
+        if (parents == null) return;
+        const src = [...parents, d.id];
+        nodes.classed("highlight", n => src.includes(n.id));
+        d3.select("#edges").selectAll("path.edgePath").classed("highlight", e => src.includes(e.v) && e.w===d.id);
+        d3.select("#edge-labels").selectAll("g.port").classed("highlight", (_, i, nodes) => {
+          const [v, w] = nodes[i].id.split("-");
+          return src.includes(v) && w===d.id;
+        });
+        e.stopPropagation();
+      });
     nodes.selectAll("rect").data(d => [d]).join("rect").attr("width", d => d.width).attr("height", d => d.height).attr("fill", d => d.color)
-      .attr("x", d => -d.width/2).attr("y", d => -d.height/2).attr("style", d => d.style ?? `stroke:#4a4b57; stroke-width:${STROKE_WIDTH}px;`);
+      .attr("x", d => -d.width/2).attr("y", d => -d.height/2).attr("class", d => d.className ?? "node");
     nodes.selectAll("g.label").data(d => [d]).join("g").attr("class", "label").attr("transform", d => {
       const x = (d.width-d.padding*2)/2;
       const y = (d.height-d.padding*2)/2+STROKE_WIDTH;
@@ -75,19 +96,19 @@ async function renderDag(graph, additions, recenter=false) {
       }
       return [ret];
     }).join("text").selectAll("tspan").data(d => d).join("tspan").attr("x", "0").attr("dy", 14).selectAll("tspan").data(d => d).join("tspan")
-      .attr("fill", d => d.color).text(d => d.st).attr("xml:space", "preserve");
+      .attr("fill", d => darkenHex(d.color, 25)).text(d => d.st).attr("xml:space", "preserve");
     addTags(nodes.selectAll("g.tag").data(d => d.tag != null ? [d] : []).join("g").attr("class", "tag")
       .attr("transform", d => `translate(${-d.width/2+8}, ${-d.height/2+8})`).datum(e => e.tag));
     // draw edges
-    const line = d3.line().x(d => d.x).y(d => d.y).curve(d3.curveBasis);
-    d3.select("#edges").selectAll("path.edgePath").data(g.edges()).join("path").attr("class", "edgePath").attr("d", (e) => {
+    const line = d3.line().x(d => d.x).y(d => d.y).curve(d3.curveBasis), edges = g.edges();
+    d3.select("#edges").selectAll("path.edgePath").data(edges).join("path").attr("class", "edgePath").attr("d", (e) => {
       const edge = g.edge(e);
       const points = edge.points.slice(1, edge.points.length-1);
       points.unshift(intersectRect(g.node(e.v), points[0]));
       points.push(intersectRect(g.node(e.w), points[points.length-1]));
       return line(points);
     }).attr("marker-end", "url(#arrowhead)");
-    addTags(d3.select("#edge-labels").selectAll("g").data(g.edges().filter(e => g.edge(e).label != null)).join("g").attr("transform", (e) => {
+    addTags(d3.select("#edge-labels").selectAll("g").data(edges).join("g").attr("transform", (e) => {
       // get a point near the end
       const [p1, p2] = g.edge(e).points.slice(-2);
       const dx = p2.x-p1.x;
@@ -101,7 +122,7 @@ async function renderDag(graph, additions, recenter=false) {
       const x = p2.x - ux * offset;
       const y = p2.y - uy * offset;
       return `translate(${x}, ${y})`
-    }).attr("class", "tag").datum(e => g.edge(e).label));
+    }).attr("class", e => g.edge(e).label.type).attr("id", e => `${e.v}-${e.w}`).datum(e => g.edge(e).label.text));
     if (recenter) document.getElementById("zoom-to-fit-btn").click();
   };
 
@@ -216,14 +237,40 @@ async function renderProfiler() {
       const peak = u64();
       const height = heightScale(peak);
       const yscale = d3.scaleLinear().domain([0, peak]).range([height, 0]);
-      const timestamps = Array.from({length:u32()}, u32);
+      let x = 0, y = 0;
+      const buf_shapes = new Map(), temp = new Map();
+      const timestamps = [];
       for (let j=0; j<eventsLen; j++) {
-        const length = u32();
-        const x = Array.from({ length }, () => timestamps[u32()]);
-        const y = Array.from({ length }, u64);
-        const dtype = strings[u32()], sz = u64(), nbytes = dtypeSize[dtype]*sz;
+        const alloc = u8(), ts = u32(), key = u32();
+        if (alloc) {
+          const dtype = strings[u32()], sz = u64(), nbytes = dtypeSize[dtype]*sz;
+          const shape = {x:[x], y:[y], dtype, sz, nbytes, key};
+          buf_shapes.set(key, shape); temp.set(key, shape);
+          timestamps.push(ts);
+          x += 1; y += nbytes;
+        } else {
+          const free = buf_shapes.get(key);
+          timestamps.push(ts);
+          x += 1; y -= free.nbytes;
+          free.x.push(x);
+          free.y.push(free.y.at(-1));
+          temp.delete(key);
+          for (const [k, v] of temp) {
+            if (k <= key) continue;
+            v.x.push(x, x);
+            v.y.push(v.y.at(-1), v.y.at(-1)-free.nbytes);
+          }
+        }
+      }
+      for (const [_, v] of temp) {
+        v.x.push(x);
+        v.y.push(v.y.at(-1));
+      }
+      timestamps.push(dur);
+      for (const [_, {dtype, sz, nbytes, y, x:steps}] of buf_shapes) {
+        const x = steps.map(s => timestamps[s]);
         const arg = {tooltipText:`${dtype} len:${formatUnit(sz)}\n${formatUnit(nbytes, "B")}`};
-        shapes.push({ x, y0:y.map(yscale), y1:y.map(y0 => yscale(y0+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, j) });
+        shapes.push({ x, y0:y.map(yscale), y1:y.map(y0 => yscale(y0+nbytes)), arg, fillColor:cycleColors(colorScheme.BUFFER, shapes.length) });
       }
       data.tracks.set(k, { shapes, offsetY, height, peak, scaleFactor:maxheight*4/height });
       div.style("height", height+padding+"px").style("cursor", "pointer").on("click", (e) => {
@@ -267,12 +314,12 @@ async function renderProfiler() {
         // generic polygon
         if (e.width == null) {
           const x = e.x.map(xscale);
-          ctx.beginPath();
-          ctx.moveTo(x[0], offsetY+e.y0[0]);
-          for (let i=1; i<x.length; i++) ctx.lineTo(x[i], offsetY+e.y0[i]);
-          for (let i=x.length-1; i>=0; i--) ctx.lineTo(x[i], offsetY+e.y1[i]);
-          ctx.closePath();
-          ctx.fill();
+          const p = new Path2D();
+          p.moveTo(x[0], offsetY+e.y0[0]);
+          for (let i=1; i<x.length; i++) p.lineTo(x[i], offsetY+e.y0[i]);
+          for (let i=x.length-1; i>=0; i--) p.lineTo(x[i], offsetY+e.y1[i]);
+          p.closePath();
+          ctx.fill(p);
           // NOTE: y coordinates are in reverse order
           for (let i = 0; i < x.length - 1; i++) {
             let tooltipText = e.arg.tooltipText;
@@ -343,8 +390,7 @@ async function renderProfiler() {
     d3.select(canvas).call(canvasZoom.transform, zoomLevel);
   }
 
-  canvasZoom = d3.zoom().filter(e => (!e.ctrlKey || e.type === 'wheel' || e.type === 'mousedown') && !e.button)
-    .scaleExtent([1, Infinity]).translateExtent([[0,0], [Infinity,0]]).on("zoom", e => render(e.transform));
+  canvasZoom = d3.zoom().filter(vizZoomFilter).scaleExtent([1, Infinity]).translateExtent([[0,0], [Infinity,0]]).on("zoom", e => render(e.transform));
   d3.select(canvas).call(canvasZoom);
   document.addEventListener("contextmenu", e => e.ctrlKey && e.preventDefault());
 
@@ -381,7 +427,8 @@ async function renderProfiler() {
 
 // ** zoom and recentering
 
-const svgZoom = d3.zoom().on("zoom", (e) => d3.select("#render").attr("transform", e.transform));
+const vizZoomFilter = e => (!e.ctrlKey || e.type === 'wheel' || e.type === 'mousedown') && !e.button && e.type !== 'dblclick';
+const svgZoom = d3.zoom().filter(vizZoomFilter).on("zoom", (e) => d3.select("#render").attr("transform", e.transform));
 d3.select("#graph-svg").call(svgZoom);
 
 // zoom to fit into view
@@ -485,7 +532,6 @@ function setState(ns) {
 
 // set a new context and keep the old one in browser history
 function setCtxWithHistory(newCtx, step=0) {
-  if (newCtx == null) return;
   // NOTE: browser does a structured clone, passing a mutable object is safe.
   history.replaceState(state, "");
   history.pushState(state, "");

tinygrad/viz/js/worker.js

@@ -8,7 +8,7 @@ onmessage = (e) => {
   const { graph, additions, ctxs } = e.data;
   const g = new dagre.graphlib.Graph({ compound: true });
   g.setGraph({ rankdir: "LR" }).setDefaultEdgeLabel(function() { return {}; });
-  if (additions.length !== 0) g.setNode("addition", {label:"", style:"fill: rgba(26, 27, 38, 0.5);", padding:0});
+  if (additions.length !== 0) g.setNode("addition", {label:"", className:"overlay", padding:0});
   for (let [k, {label, src, ref, ...rest }] of Object.entries(graph)) {
     // adjust node dims by label size (excluding escape codes) + add padding
     let [width, height] = [0, 0];
@@ -16,11 +16,11 @@ onmessage = (e) => {
       width = Math.max(width, ctx.measureText(line).width);
       height += LINE_HEIGHT;
     }
-    g.setNode(k, {width:width+NODE_PADDING*2, height:height+NODE_PADDING*2, padding:NODE_PADDING, label, ref, ...rest});
+    g.setNode(k, {width:width+NODE_PADDING*2, height:height+NODE_PADDING*2, padding:NODE_PADDING, label, ref, id:k, ...rest});
     // add edges
     const edgeCounts = {}
-    for (const s of src) edgeCounts[s] = (edgeCounts[s] || 0)+1;
-    for (const s of src) g.setEdge(s, k, { label: edgeCounts[s] > 1 ? edgeCounts[s] : null });
+    for (const [_, s] of src) edgeCounts[s] = (edgeCounts[s] || 0)+1;
+    for (const [port, s] of src) g.setEdge(s, k, { label: edgeCounts[s] > 1 ? {type:"tag", text:edgeCounts[s]} : {type:"port", text:port}});
     if (additions.includes(parseInt(k))) g.setParent(k, "addition");
   }
   dagre.layout(g);

tinygrad/viz/serve.py

@@ -11,6 +11,7 @@ from tinygrad.uop.ops import TrackedGraphRewrite, UOp, Ops, printable, GroupOp,
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent, ProfileGraphEntry, Device
 from tinygrad.renderer import ProgramSpec
 from tinygrad.dtype import dtypes
+from tinygrad.codegen.opt.kernel import axis_colors
 
 uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0", Ops.VCONST: "#e0e0e0", Ops.REDUCE: "#FF5B5B",
                Ops.DEFINE_GLOBAL: "#ffe0b0", Ops.DEFINE_LOCAL: "#ffe0d0", Ops.DEFINE_REG: "#f0ffe0", Ops.REDUCE_AXIS: "#FF6B6B",
@@ -79,13 +80,13 @@ def uop_to_json(x:UOp) -> dict[int, dict]:
       if u.op not in {Ops.VIEW, Ops.BUFFER, Ops.KERNEL, Ops.ASSIGN, Ops.COPY, Ops.SINK, *GroupOp.Buffer} and u.st is not None:
         label += f"\n{shape_to_str(u.shape)}"
       elif len(rngs:=u.ranges):
-        label += f"\n{str(sorted([x.arg[0] for x in rngs]))}"
+        label += f"\n({','.join([colored(str(x.arg[0]), axis_colors[x.arg[-1]]) for x in sorted(rngs, key=lambda x: x.arg[0:-1])])})"
     except Exception:
       label += "\n<ISSUE GETTING LABEL>"
     if (ref:=ref_map.get(u.arg.ast) if u.op is Ops.KERNEL else None) is not None: label += f"\ncodegen@{ctxs[ref]['name']}"
     # NOTE: kernel already has metadata in arg
     if TRACEMETA >= 2 and u.metadata is not None and u.op is not Ops.KERNEL: label += "\n"+repr(u.metadata)
-    graph[id(u)] = {"label":label, "src":[id(x) for x in u.src if x not in excluded], "color":uops_colors.get(u.op, "#ffffff"),
+    graph[id(u)] = {"label":label, "src":[(i,id(x)) for i,x in enumerate(u.src) if x not in excluded], "color":uops_colors.get(u.op, "#ffffff"),
                     "ref":ref, "tag":u.tag}
   return graph
 
@@ -154,37 +155,22 @@ def timeline_layout(dev_events:list[tuple[int, int, float, DevEvent]], start_ts:
 
 def mem_layout(events:list[tuple[int, int, float, DevEvent]], start_ts:int, end_ts:int, peaks:list[int], dtype_size:dict[str, int],
                scache:dict[str, int]) -> bytes|None:
-  step, peak, mem = 0, 0, 0
-  shps:dict[int, dict] = {}
-  temp:dict[int, dict] = {}
-  timestamps:list[int] = []
+  peak, mem = 0, 0
+  temp:dict[int, int] = {}
+  bufs:list[bytes] = []
   for st,_,_,e in events:
     if not isinstance(e, ProfilePointEvent): continue
     if e.name == "alloc":
-      shps[e.key] = temp[e.key] = {"x":[step], "y":[mem], "arg":{"dtype":e.arg["dtype"].name, "sz":e.arg["sz"]}}
+      bufs.append(struct.pack("<BIIIQ", 1, int(e.ts)-start_ts, e.key, enum_str(e.arg["dtype"].name, scache), e.arg["sz"]))
       dtype_size.setdefault(e.arg["dtype"].name, e.arg["dtype"].itemsize)
-      timestamps.append(int(e.ts)-start_ts)
-      step += 1
-      mem += e.arg["sz"]*e.arg["dtype"].itemsize
+      temp[e.key] = nbytes = e.arg["sz"]*e.arg["dtype"].itemsize
+      mem += nbytes
       if mem > peak: peak = mem
     if e.name == "free":
-      timestamps.append(int(e.ts)-start_ts)
-      step += 1
-      mem -= (free_nbytes:=(removed:=temp.pop(e.key))["arg"]["sz"]*dtype_size[removed["arg"]["dtype"]])
-      removed["x"].append(step)
-      removed["y"].append(removed["y"][-1])
-      for k,v in temp.items():
-        if k > e.key:
-          v["x"] += [step, step]
-          v["y"] += [v["y"][-1], v["y"][-1]-free_nbytes]
-  for v in temp.values():
-    v["x"].append(step)
-    v["y"].append(v["y"][-1])
-  timestamps.append(end_ts-start_ts)
+      bufs.append(struct.pack("<BII", 0, int(e.ts)-start_ts, e.key))
+      mem -= temp.pop(e.key)
   peaks.append(peak)
-  bufs = [struct.pack("<I"+str(i:=len(v['x']))+f"I{i}QIQ", i, *v["x"], *v["y"], enum_str(v["arg"]["dtype"], scache),
-                      v["arg"]["sz"]) for v in shps.values()]
-  return struct.pack("<BIQI", 1, len(shps), peak, len(timestamps))+struct.pack(f"<{len(timestamps)}I", *timestamps)+b"".join(bufs) if bufs else None
+  return struct.pack("<BIQ", 1, len(bufs), peak)+b"".join(bufs) if bufs else None
 
 def get_profile(profile:list[ProfileEvent]) -> bytes|None:
   # start by getting the time diffs
@@ -272,7 +258,7 @@ class Handler(BaseHTTPRequestHandler):
       if url.path == "/disasm": ret, content_type = get_disassembly(**query), "application/json"
       else: return self.stream_json(get_details(contexts[1][int(query["ctx"][0])][int(query["idx"][0])]))
     elif url.path == "/ctxs": ret, content_type = json.dumps(ctxs).encode(), "application/json"
-    elif url.path == "/get_profile" and profile_ret is not None: ret, content_type = profile_ret, "application/json"
+    elif url.path == "/get_profile" and profile_ret: ret, content_type = profile_ret, "application/octet-stream"
     else: status_code = 404
 
     # send response
@@ -305,8 +291,8 @@ def reloader():
       os.execv(sys.executable, [sys.executable] + sys.argv)
     time.sleep(0.1)
 
-def load_pickle(path:str):
-  if path is None or not os.path.exists(path): return None
+def load_pickle(path:str|None) -> list:
+  if path is None or not os.path.exists(path): return []
   with open(path, "rb") as f: return pickle.load(f)
 
 # NOTE: using HTTPServer forces a potentially slow socket.getfqdn
@@ -329,16 +315,16 @@ if __name__ == "__main__":
   contexts, profile = load_pickle(args.kernels), load_pickle(args.profile)
 
   # NOTE: this context is a tuple of list[keys] and list[values]
-  ctxs = get_metadata(*contexts[:2]) if contexts is not None else []
+  ctxs = get_metadata(*contexts[:2]) if contexts else []
 
-  profile_ret = get_profile(profile) if profile is not None else None
+  profile_ret = get_profile(profile)
 
   server = TCPServerWithReuse(('', PORT), Handler)
   reloader_thread = threading.Thread(target=reloader)
   reloader_thread.start()
   print(f"*** started viz on {HOST}:{PORT}")
   print(colored(f"*** ready in {(time.perf_counter()-st)*1e3:4.2f}ms", "green"), flush=True)
-  if len(getenv("BROWSER", "")) > 0: webbrowser.open(f"{HOST}:{PORT}{'/profiler' if contexts is None else ''}")
+  if len(getenv("BROWSER", "")) > 0: webbrowser.open(f"{HOST}:{PORT}")
   try: server.serve_forever()
   except KeyboardInterrupt:
     print("*** viz is shutting down...")