tilelang-rocm 0.1.4.post10__cp310-cp310-manylinux1_x86_64.whl → 0.1.4.post12__cp310-cp310-manylinux1_x86_64.whl

tilelang/VERSION

@@ -1 +1 @@
-0.1.4.post10
+0.1.4.post12

tilelang/autotuner/__init__.py

@@ -167,7 +167,7 @@ class AutoTuner:
                          max_mismatched_ratio: float = 0.01,
                          skip_check: bool = False,
                          manual_check_prog: Callable = None,
-                         cache_input_tensors: bool = True):
+                         cache_input_tensors: bool = False):
         """Set profiling arguments for the auto-tuner.
 
         Args:
@@ -463,8 +463,26 @@ class _AutoTunerImplementation:
     rep: int = 100
     timeout: int = 100
     configs: Any = None
-
-    def __init__(self, configs: Any, warmup: int = 25, rep: int = 100, timeout: int = 100) -> None:
+    supply_type: tilelang.TensorSupplyType = tilelang.TensorSupplyType.Auto
+    ref_prog: Callable = None
+    supply_prog: Callable = None
+    rtol: float = 1e-2
+    atol: float = 1e-2
+    max_mismatched_ratio: float = 0.01
+    skip_check: bool = False
+    manual_check_prog: Callable = None
+    cache_input_tensors: bool = False
+
+    def __init__(self, configs: Any, warmup: int = 25, rep: int = 100, timeout: int = 100,
+                 supply_type: tilelang.TensorSupplyType = tilelang.TensorSupplyType.Auto,
+                 ref_prog: Callable = None,
+                 supply_prog: Callable = None,
+                 rtol: float = 1e-2,
+                 atol: float = 1e-2,
+                 max_mismatched_ratio: float = 0.01,
+                 skip_check: bool = False,
+                 manual_check_prog: Callable = None,
+                 cache_input_tensors: bool = False) -> None:
         """Initialize the AutoTunerImplementation.
 
         Args:
@@ -509,7 +527,17 @@ class _AutoTunerImplementation:
                 def jit_compile(**config_arg):
                     return fn(*args, **kwargs, __tune_params=config_arg)
 
-                autotuner = AutoTuner(fn, configs=configs)
+                autotuner = AutoTuner(fn, configs=configs).set_profile_args(
+                    supply_type=self.supply_type,
+                    ref_prog=self.ref_prog,
+                    supply_prog=self.supply_prog,
+                    rtol=self.rtol,
+                    atol=self.atol,
+                    max_mismatched_ratio=self.max_mismatched_ratio,
+                    skip_check=self.skip_check,
+                    manual_check_prog=self.manual_check_prog,
+                    cache_input_tensors=self.cache_input_tensors,
+                )
                 autotuner.jit_compile = jit_compile
                 autotuner.run = partial(autotuner.run, warmup, rep, timeout)
 
@@ -525,9 +553,21 @@ def autotune(  # This is the new public interface
         func: Union[Callable[_P, _RProg], PrimFunc, None] = None,
         *,  # Indicates subsequent arguments are keyword-only
         configs: Any,
+        # profile arguments
         warmup: int = 25,
         rep: int = 100,
-        timeout: int = 100):
+        timeout: int = 100,
+        # compile arguments
+        supply_type: tilelang.TensorSupplyType = tilelang.TensorSupplyType.Auto,
+        ref_prog: Callable = None,
+        supply_prog: Callable = None,
+        rtol: float = 1e-2,
+        atol: float = 1e-2,
+        max_mismatched_ratio: float = 0.01,
+        skip_check: bool = False,
+        manual_check_prog: Callable = None,
+        cache_input_tensors: bool = False,
+    ):
     """
     Just-In-Time (JIT) compiler decorator for TileLang functions.
 
@@ -571,5 +611,15 @@ def autotune(  # This is the new public interface
         # Create a _AutoTunerImplementation instance with the provided/defaulted arguments.
         # This instance is a decorator that will be applied to the function later.
         configured_decorator = _AutoTunerImplementation(
-            configs=configs, warmup=warmup, rep=rep, timeout=timeout)
+            configs=configs, warmup=warmup, rep=rep, timeout=timeout,
+            supply_type=supply_type,
+            ref_prog=ref_prog,
+            supply_prog=supply_prog,
+            rtol=rtol,
+            atol=atol,
+            max_mismatched_ratio=max_mismatched_ratio,
+            skip_check=skip_check,
+            manual_check_prog=manual_check_prog,
+            cache_input_tensors=cache_input_tensors,
+        )
         return configured_decorator

tilelang/engine/phase.py

@@ -142,13 +142,14 @@ def OptimizeForTarget(mod: IRModule, target: Target) -> IRModule:
     mod = tilelang.transform.AnnotateDeviceRegions()(mod)
     mod = tir.transform.SplitHostDevice()(mod)
 
-    if allow_warp_specialized(pass_ctx=pass_ctx, target=target):
-        # This is a workaround to avoid the bug in the MergeSharedMemoryAllocations pass
-        # when warp specialization is enabled, as different warp threads may access different
-        # buffers, but the liveness analysis is hard because we need to do pipeline.
-        mod = tir.transform.MergeSharedMemoryAllocations()(mod)
-    else:
-        mod = tilelang.transform.MergeSharedMemoryAllocations()(mod)
+    mod = tir.transform.MergeSharedMemoryAllocations()(mod)
+    # if allow_warp_specialized(pass_ctx=pass_ctx, target=target):
+    #     # This is a workaround to avoid the bug in the MergeSharedMemoryAllocations pass
+    #     # when warp specialization is enabled, as different warp threads may access different
+    #     # buffers, but the liveness analysis is hard because we need to do pipeline.
+    #     mod = tir.transform.MergeSharedMemoryAllocations()(mod)
+    # else:
+    #     mod = tilelang.transform.MergeSharedMemoryAllocations()(mod)
 
     mod = tilelang.transform.ThreadSync("shared")(mod)
     mod = tilelang.transform.ThreadSync("shared.dyn")(mod)

tilelang/intrinsics/mfma_macro_generator.py

@@ -358,6 +358,8 @@ class MatrixCoreIntrinEmitter(object):
         BLOCK_M = block_row_warps * warp_rows
         BLOCK_N = block_col_warps * warp_cols
         M_DIM, N_DIM = self.M_DIM, self.N_DIM
+        C_buf_dims = len(C_buf.shape)
+        assert C_buf_dims in {2, 4}, "C_buf should be 2D or 4D"
 
         # STS
         # MMA Store must be in simulated instead of TVM Intrins
@@ -369,9 +371,15 @@ class MatrixCoreIntrinEmitter(object):
             for i, j in T.grid(warp_rows, warp_cols):
                 for local_id in T.vectorized(local_size_out):
                     row, col = T.meta_var(mfma_store_index_map(tx, local_id))
-                    C_buf[warp_m * warp_rows + i, warp_n * warp_cols + j, row,
-                          col] = C_local_buf[i * warp_cols * local_size_out + j * local_size_out +
-                                             local_id]
+                    if C_buf_dims == 2:
+                        C_buf[(warp_m * warp_rows + i) * M_DIM + row,
+                                  (warp_n * warp_cols + j) * N_DIM +
+                                  col] = C_local_buf[i * (warp_cols * local_size_out) +
+                                                     j * local_size_out + local_id]
+                    else:
+                        C_buf[warp_m * warp_rows + i, warp_n * warp_cols + j, row,
+                            col] = C_local_buf[i * warp_cols * local_size_out + j * local_size_out +
+                                                local_id]
 
         @T.macro
         def _warp_stmatrix_global(C_local_buf, C_buf, thread_binding):

tilelang/jit/adapter/cython/adapter.py

@@ -432,8 +432,8 @@ class CythonKernelAdapter(BaseKernelAdapter):
     def _convert_torch_func(self) -> Callable:
         """Returns a PyTorch-compatible function wrapper for the kernel."""
 
-        def lambda_forward(*args, stream: int = -1):
-            return self.cython_wrapper.forward([*args], stream=stream)
+        def lambda_forward(*args, stream: int = -1, skip_check: bool = False):
+            return self.cython_wrapper.forward([*args], stream=stream, skip_check=skip_check)
 
         return lambda_forward
 

tilelang/jit/adapter/cython/cython_wrapper.pyx

@@ -66,7 +66,7 @@ cdef class CythonKernelWrapper:
         self.buffer_device_map = buffer_device_map
         return self
 
-    cpdef forward(self, list inputs, int64_t stream = -1):
+    cpdef forward(self, list inputs, int64_t stream = -1, bint skip_check = False):
         # Validate input dimensions and prepare for kernel execution
         cdef int total_params = len(self.params)
         cdef int total_inputs = len(inputs)
@@ -135,29 +135,30 @@ cdef class CythonKernelWrapper:
                 raise ValueError(f"Unsupported tensor type: {type(tensor)}")
 
         # Check buffer device
-        # cdef str tensor_list_device_type = tensor_list[0].device.type
-        if isinstance(tensor_list[0], torch.Tensor):
-            tensor_list_device_type = tensor_list[0].device.type
-        for param, (buffer_idx, device) in self.buffer_device_map.items():
-            if isinstance(tensor_list[buffer_idx], torch.Tensor):
-                tensor_device = tensor_list[buffer_idx].device
-                # Compare device types and indices separately to handle both string and torch.device objects            
-                if (tensor_list_device_type != device.type or 
-                    (tensor_device.index is not None and device.index is not None and tensor_device.index != device.index)):
-                    raise ValueError(f"Buffer device mismatch for parameter {param}: expected {device}, got {tensor_device}")
-
-        # Check buffer dtype map
-        for param, (buffer_idx, torch_dtype) in self.buffer_dtype_map.items():
-            if isinstance(tensor_list[buffer_idx], torch.Tensor):
-                if tensor_list[buffer_idx].dtype != torch_dtype:
-                    raise ValueError(f"Buffer dtype mismatch for parameter {param}: expected {torch_dtype}, got {tensor_list[buffer_idx].dtype}")
-        
-        # Check static shape map
-        for param, (buffer_idx, shape_list) in self.static_shape_map.items():
-            if isinstance(tensor_list[buffer_idx], torch.Tensor):
-                for shape_idx, shape in shape_list:
-                    if tensor_list[buffer_idx].shape[shape_idx] != shape:
-                        raise ValueError(f"Static shape mismatch for parameter {param}: expected {shape} at index {shape_idx}, got {tensor_list[buffer_idx].shape}")
+        if not skip_check:
+            # cdef str tensor_list_device_type = tensor_list[0].device.type
+            if isinstance(tensor_list[0], torch.Tensor):
+                tensor_list_device_type = tensor_list[0].device.type
+            for param, (buffer_idx, device) in self.buffer_device_map.items():
+                if isinstance(tensor_list[buffer_idx], torch.Tensor):
+                    tensor_device = tensor_list[buffer_idx].device
+                    # Compare device types and indices separately to handle both string and torch.device objects            
+                    if (tensor_list_device_type != device.type or 
+                        (tensor_device.index is not None and device.index is not None and tensor_device.index != device.index)):
+                        raise ValueError(f"Buffer device mismatch for parameter {param}: expected {device}, got {tensor_device}")
+
+            # Check buffer dtype map
+            for param, (buffer_idx, torch_dtype) in self.buffer_dtype_map.items():
+                if isinstance(tensor_list[buffer_idx], torch.Tensor):
+                    if tensor_list[buffer_idx].dtype != torch_dtype:
+                        raise ValueError(f"Buffer dtype mismatch for parameter {param}: expected {torch_dtype}, got {tensor_list[buffer_idx].dtype}")
+            
+            # Check static shape map
+            for param, (buffer_idx, shape_list) in self.static_shape_map.items():
+                if isinstance(tensor_list[buffer_idx], torch.Tensor):
+                    for shape_idx, shape in shape_list:
+                        if tensor_list[buffer_idx].shape[shape_idx] != shape:
+                            raise ValueError(f"Static shape mismatch for parameter {param}: expected {shape} at index {shape_idx}, got {tensor_list[buffer_idx].shape}")
 
         # Add dynamic dimension values to kernel arguments
         for _, (buffer_idx, shape_idx) in self.dynamic_symbolic_map.items():

tilelang/src/tl_templates/hip/common.h

@@ -99,3 +99,16 @@ TL_DEVICE unsigned __pack_half2(const half_t x, const half_t y) {
   unsigned v1 = *((unsigned short *)&y);
   return (v1 << 16) | v0;
 }
+
+// Pack two bfloat16_t values.
+TL_DEVICE unsigned __pack_bfloat162(const bfloat16_t x, const bfloat16_t y) {
+  unsigned v0 = *((unsigned short *)&x);
+  unsigned v1 = *((unsigned short *)&y);
+  return (v1 << 16) | v0;
+}
+
+
+template <typename T1, typename T2>
+TL_DEVICE void AtomicAdd(T1 *address, T2 val) {
+  atomicAdd(reinterpret_cast<T1 *>(address), static_cast<T1>(val));
+}

tilelang/src/tl_templates/hip/hip_fp8.h

@@ -16,3 +16,33 @@ struct __align__(16) fp8_e4_16_t {
   fp8_e4_8_t x;
   fp8_e4_8_t y;
 };
+
+__device__ fp8_e4_4_t make_fp8_e4_4_t(fp8_e4_t x, fp8_e4_t y, fp8_e4_t z,
+                                      fp8_e4_t w) {
+  // reinterpret the 4 fp8_e4_t values to signed char value and shift
+  signed char x_char = *reinterpret_cast<signed char *>(&x);
+  signed char y_char = *reinterpret_cast<signed char *>(&y);
+  signed char z_char = *reinterpret_cast<signed char *>(&z);
+  signed char w_char = *reinterpret_cast<signed char *>(&w);
+  int res = (w_char << 24) | (z_char << 16) | (y_char << 8) | x_char;
+  return *reinterpret_cast<fp8_e4_4_t *>(&res);
+}
+
+__device__ fp8_e4_8_t make_fp8_e4_8_t(fp8_e4_t x, fp8_e4_t y, fp8_e4_t z,
+                                      fp8_e4_t w, fp8_e4_t v, fp8_e4_t u,
+                                      fp8_e4_t t, fp8_e4_t s) {
+  signed char x_char = *reinterpret_cast<signed char *>(&x);
+  signed char y_char = *reinterpret_cast<signed char *>(&y);
+  signed char z_char = *reinterpret_cast<signed char *>(&z);
+  signed char w_char = *reinterpret_cast<signed char *>(&w);
+  signed char v_char = *reinterpret_cast<signed char *>(&v);
+  signed char u_char = *reinterpret_cast<signed char *>(&u);
+  signed char t_char = *reinterpret_cast<signed char *>(&t);
+  signed char s_char = *reinterpret_cast<signed char *>(&s);
+  int a = (w_char << 24) | (z_char << 16) | (y_char << 8) | x_char;
+  int b = (s_char << 24) | (t_char << 16) | (u_char << 8) | v_char;
+  fp8_e4_8_t res;
+  res.x = *reinterpret_cast<fp8_e4_4_t *>(&a);
+  res.y = *reinterpret_cast<fp8_e4_4_t *>(&b);
+  return res;
+}

{tilelang_rocm-0.1.4.post10.dist-info → tilelang_rocm-0.1.4.post12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tilelang-rocm
-Version: 0.1.4.post10
+Version: 0.1.4.post12
 Summary: A tile level programming language to generate high performance code.
 Home-page: https://github.com/tile-ai/tilelang
 Author: Microsoft Research

{tilelang_rocm-0.1.4.post10.dist-info → tilelang_rocm-0.1.4.post12.dist-info}/RECORD

@@ -1,7 +1,7 @@
 tilelang/CMakeLists.txt,sha256=xJhnusYZI4UhD_fzseGH3Tn2BeovUzz3aWUwPq-WU0Y,7010
 tilelang/LICENSE,sha256=v9fVeAgRKQXc5ySwTns767gj0-dHN9XYPpGURkAVAXs,1127
 tilelang/README.md,sha256=1RC_2IUBY-p0BR-d2xkNXC8zrva8-U3AVkmCozkssbY,11924
-tilelang/VERSION,sha256=Il_uPi-DubG_kcKlTV5qP75Dq4zOcrugFwCAygQ4SQU,13
+tilelang/VERSION,sha256=Axg99nX6i4uA5L-33bP_V3uMQL8--B3WbQ83pCXrIfg,13
 tilelang/__init__.py,sha256=yH0BknCRnFQN-E7d6p1HPNbeY4o3COqG7XzR_EJpbTo,3215
 tilelang/_ffi_api.py,sha256=D-HfDxx8EZq6qItftg-ejOhpC_smIZLN-pWPVCNX_UM,243
 tilelang/config.cmake,sha256=370i6N3wwi7-LPGZDBtiiu54UWp39ndD-9lCurLhHwI,14330
@@ -6280,7 +6280,7 @@ tilelang/3rdparty/tvm/src/topi/reduction.cc,sha256=vjnvD9lnmpO57_CNXLizVnimXRffm
 tilelang/3rdparty/tvm/src/topi/schedule.cc,sha256=pazCd8PkTy57jksqinRZVPvJ18iciI0dJXRA0YRACqI,12522
 tilelang/3rdparty/tvm/src/topi/transform.cc,sha256=j9doFFms3ba_1r-Wt-t9Wq953ze-DJeO9uskDtlsejc,7862
 tilelang/3rdparty/tvm/src/topi/vision.cc,sha256=1c0j7VN6orhaRF980GZY8ynLOT66Jq8SFk0GIEk-nrw,1245
-tilelang/autotuner/__init__.py,sha256=7c5xFLUtqUjFUNmmxutr5G69Eotn51dQuDtvhlC1-ks,22446
+tilelang/autotuner/__init__.py,sha256=eCbbYp0Gz_o8Yinq5N6xV8y4siQ59ay0b3Ilq773i4I,24519
 tilelang/autotuner/param.py,sha256=e_vRYHXjKHEsloJ0vwhC4Peqa2vN640DFS3bkQyCrfQ,11802
 tilelang/cache/__init__.py,sha256=ZNBkLL3ssxtMMq8LFTThreL4fSGYbiVLcyD7YH5bb8U,1869
 tilelang/cache/kernel_cache.py,sha256=VZxu3vnFML-Der37ujZlAgapSitQvFYWeixBgIQrGms,13129
@@ -6329,10 +6329,10 @@ tilelang/engine/__init__.py,sha256=8veNHGeKyssdX_3ytZEsu0jKHPlP5ygBMUhXXTkeUHc,2
 tilelang/engine/callback.py,sha256=zB_lp86j7jVPXnwSg0LMMWltxmkJQiU0wlxFWK-sxf4,3275
 tilelang/engine/lower.py,sha256=OJX7d_qk6WXA9VDF-fTTKw1q-ZnBkn4tuD9lA4TPdxk,8923
 tilelang/engine/param.py,sha256=5eWc48aao84WIrbtaLuAYGrb3RE1SyiG6hIy8DwlryI,3943
-tilelang/engine/phase.py,sha256=SPruc1tsC_yFD9q93pilkFmzY3UZ2YJTCRH06YPDI1Q,7154
+tilelang/engine/phase.py,sha256=vt0BzpkYDsXfSHH2BPqSfGQwhiF6TWm9HwgQVKsr1DQ,7228
 tilelang/intrinsics/__init__.py,sha256=ymvtsKjVY0f_9k-QIMtO4CEh6hEnG7H4NiW3buNLVQg,501
 tilelang/intrinsics/mfma_layout.py,sha256=O5jh8gfILH-ASDss0pvTZhJq2jgmcyHwRPzigxGoGW8,4056
-tilelang/intrinsics/mfma_macro_generator.py,sha256=axkPrTN-Lplt0xFMk-ckQu8Qynve-lNl6thQCIwY1Q0,17653
+tilelang/intrinsics/mfma_macro_generator.py,sha256=FMTAgq249F8APp8Ms8rAU-PPCX5Att_9Hu76ZyIvE64,18141
 tilelang/intrinsics/mma_layout.py,sha256=eHFiNKd3zKzNFuRrpZdEQx0apbHWj8Ak6Q3e9_CeDiM,5090
 tilelang/intrinsics/mma_macro_generator.py,sha256=BVx3Bt3K67XSTp-Op425OrPxmtD43jyLre0wY8AeW-w,44478
 tilelang/intrinsics/utils.py,sha256=dbQpWOy0F4rg3WotzHQToPtJgY2BLtRy1CKkSnrs--k,4243
@@ -6349,8 +6349,8 @@ tilelang/jit/adapter/wrapper.py,sha256=MJsMJxR-Lg9XLNkgMV_RbhQIuPPC0QAOc94H6JIkM
 tilelang/jit/adapter/ctypes/__init__.py,sha256=WA38dJGWPXQbOu_rnU1wgckOiVTMzzvxGEL7x1naYM0,127
 tilelang/jit/adapter/ctypes/adapter.py,sha256=mpqlT2Qh0WPUfWuvdO6fFAAXF2mfqbKUcltCU-M-apE,11365
 tilelang/jit/adapter/cython/__init__.py,sha256=v-6kWB8ktC1MUzsCkyYsDM34asotUIGXopEco5nqUVE,127
-tilelang/jit/adapter/cython/adapter.py,sha256=6XNStHYZOANtrGTqh3o0N6-FJbdyiEfCdtQmV3q4-aQ,19572
-tilelang/jit/adapter/cython/cython_wrapper.pyx,sha256=d8tp6MIed4BpiQHLsiDlX4P1GJ2yfrQ8bDnMR9CLXVc,8598
+tilelang/jit/adapter/cython/adapter.py,sha256=m_mRDIzWnE-lXlX29KYUiOtbLUvut4ctZU5nlQ0vKoU,19621
+tilelang/jit/adapter/cython/cython_wrapper.pyx,sha256=HO704otcqNIV5Q4Et9jyfS4WLK5QwGlyiLv6aZJMbzU,8738
 tilelang/language/__init__.py,sha256=xnBuLhX29CjAqo6B1fiF1ctWTDjYPhW-wim9R0AxVFc,5906
 tilelang/language/allocate.py,sha256=YeZERWoqdZXaYbw58hoJ61VxrD3zsiI1sJIyg6vNy38,2855
 tilelang/language/builtin.py,sha256=sLg88K7qkI3jP2ks6TR3YULlCiU9f_cWJ9c79Pi_M5w,9959
@@ -6383,10 +6383,10 @@ tilelang/layout/__init__.py,sha256=F1wr9yBG9GW84h8KWXz-hRJFfqyZuY0EKSrG08KyrWQ,2
 tilelang/layout/fragment.py,sha256=zTv9P96lsYi9BWc5pxR4PA2Z5RSDGP7D5uJCiNw7_oc,8445
 tilelang/layout/layout.py,sha256=20CWxz_S8k_WNvWiR4gdIrEsQ36e5bsnOEqmu4zGk_c,4311
 tilelang/layout/swizzle.py,sha256=PMqu_s1sNCh9uo8eDs5qmLKXnDqZwv34GT3H9D4YDO0,438
-tilelang/lib/libtilelang.so,sha256=gWKpADoERYHrDu-eOXjYtG8ak-dvv_2cs8WDiFYnc8o,5040880
-tilelang/lib/libtilelang_module.so,sha256=eStBjzb0e9y3lZVc21EefE1_nryt0Cs0wlYyz32QifI,5040880
-tilelang/lib/libtvm.so,sha256=7qy1XeUSGm4YXJC22LUmLZMfaGEPdTCXVxq3pkqS_ZY,83982112
-tilelang/lib/libtvm_runtime.so,sha256=BgUg1x-f2q3urmCbmBrbUtVFmS3SUyI6dBSAVP_yi8k,4794680
+tilelang/lib/libtilelang.so,sha256=ZKzH7wCpSuIPThuUIfB1brqjN6lE4IGKTNqfJJNwliI,5040720
+tilelang/lib/libtilelang_module.so,sha256=E23iuTGfHqKQOq8mp-L7Gx91zeCnXZLRC-wZhUNDqUs,5040720
+tilelang/lib/libtvm.so,sha256=TMs6mrRyGE9_HJ8ZoxE_gBrOgmQX_VplIu0fXeIZ9-o,83982112
+tilelang/lib/libtvm_runtime.so,sha256=a2o2Zu3-6wPZZXlcEVeCIjIdJYihgv4fw26aLWnLQ8Y,4794680
 tilelang/math/__init__.py,sha256=JC4fqrU_LV_wDErti-wHNr4j6_mqP1PsK0qqkhaSzRU,209
 tilelang/primitives/__init__.py,sha256=10gQN3QWUFM1nkGXY46QFcWUXxwsKMsVn23JdyFHil4,167
 tilelang/primitives/gemm/__init__.py,sha256=j62ObmbL5Q6m3lSouNBQDk1hZZRnSp4UNNCCaSlKYXU,1658
@@ -6416,11 +6416,11 @@ tilelang/src/tl_templates/cuda/gemm_sm90.h,sha256=S3v63snxR_3TEg8LLCoR6cqGVgii8k
 tilelang/src/tl_templates/cuda/ldsm.h,sha256=TxCxYVzUK4tvUNVqULCL5HEaAuW9vOv0_-QYmoRFUkM,5053
 tilelang/src/tl_templates/cuda/reduce.h,sha256=U9mKcHSttin1FQ0BohpaP0bHvgPvb3t-czwTuDeK5-8,4394
 tilelang/src/tl_templates/cuda/threadblock_swizzle.h,sha256=GIXQwC1gzwUhnq4CzORHh5hA_QHVfMrOcUeGTy1Fon8,1945
-tilelang/src/tl_templates/hip/common.h,sha256=KzMntW2OlABMwYjlxNqKiBDZQXQNrYApeVf8F9548s4,3591
+tilelang/src/tl_templates/hip/common.h,sha256=C4O6p1bStNX3hmvZm8QLiWJphTf-4a8OirZmOEnBhBk,3967
 tilelang/src/tl_templates/hip/copy.h,sha256=fGHkbe4ReXoEtIWrgQ-mlCycaIL65SvNGWK1OJZdUQo,3324
 tilelang/src/tl_templates/hip/debug.h,sha256=9xGr4ka5x_nvY55XwbgTJFFwEnd09ta9jAZwjHyQau0,8231
 tilelang/src/tl_templates/hip/gemm.h,sha256=lYeOjV8OG2oZbcS7ByzOudE7i0FQJ71mrUcImkfhTrg,11610
-tilelang/src/tl_templates/hip/hip_fp8.h,sha256=JYGiuuroLQH7CXT7IdKcpNUECmGOTe8DjIjcS9eLc0U,377
+tilelang/src/tl_templates/hip/hip_fp8.h,sha256=eFYV5OyHI2X5xMNOjv1NLqR2shlKkhiGQWiVEiY6LPs,1903
 tilelang/src/tl_templates/hip/ldsm.h,sha256=gRx_bSdsCsgcVumwUJwOnv4HuHruU2kC9TE9x_jo8k0,106
 tilelang/src/tl_templates/hip/reduce.h,sha256=-VKpG-TNbzPHIqsSReYpqZoM-oXFzIx6fMeBieV26Kc,1372
 tilelang/src/tl_templates/hip/threadblock_swizzle.h,sha256=HPuOUl2Ts76ckY5kVkLlfaK2WyHY0pEXskbXnnefFcA,2055
@@ -6437,8 +6437,8 @@ tilelang/utils/deprecated.py,sha256=CiZ9y_76_dZ24SFDdasDiLmibwi6xO2Gdj6WzTWU0Qg,
 tilelang/utils/language.py,sha256=KUzUZ8Z2x1np0Hu_MrjWOIcRrVAZHX90li1Xw9fYZXY,3291
 tilelang/utils/target.py,sha256=P-74pdCLWcp2MZMQUoPIFwKF1NZ1QT-L0VroIL8m2to,2486
 tilelang/utils/tensor.py,sha256=SZ4ewoJ-Mq3zg8zIHS7-XLUmYDdlNwh841yUkjnQtNU,12573
-tilelang_rocm-0.1.4.post10.dist-info/licenses/LICENSE,sha256=v9fVeAgRKQXc5ySwTns767gj0-dHN9XYPpGURkAVAXs,1127
-tilelang_rocm-0.1.4.post10.dist-info/METADATA,sha256=aBSwulXbZ0_bU2cIYKNCuRsUM2qlAsJ6O14B7fPJUoQ,13076
-tilelang_rocm-0.1.4.post10.dist-info/WHEEL,sha256=0-G7woG4LgutcYzUGJCOYFgoh749-FtfhSMeIPLVGS0,104
-tilelang_rocm-0.1.4.post10.dist-info/top_level.txt,sha256=qvMq-AYkDVggI-9VIAzCe5CXHl66IEWj7J29-JbuFsI,21
-tilelang_rocm-0.1.4.post10.dist-info/RECORD,,
+tilelang_rocm-0.1.4.post12.dist-info/licenses/LICENSE,sha256=v9fVeAgRKQXc5ySwTns767gj0-dHN9XYPpGURkAVAXs,1127
+tilelang_rocm-0.1.4.post12.dist-info/METADATA,sha256=NTJMFhxd7258jXSi3h5eBbJNpO0HpAfXviyafU8ouvQ,13076
+tilelang_rocm-0.1.4.post12.dist-info/WHEEL,sha256=0-G7woG4LgutcYzUGJCOYFgoh749-FtfhSMeIPLVGS0,104
+tilelang_rocm-0.1.4.post12.dist-info/top_level.txt,sha256=qvMq-AYkDVggI-9VIAzCe5CXHl66IEWj7J29-JbuFsI,21
+tilelang_rocm-0.1.4.post12.dist-info/RECORD,,