You are viewing a plain text version of this content. The canonical link for it is here.
Posted to discuss-archive@tvm.apache.org by "akaashrp (via GitHub)" <gi...@apache.org> on 2026/08/22 07:27:26 UTC

[PR] [Web] Decode packed BF16 tensor records in place [tvm]

akaashrp opened a new pull request, #20167:
URL: https://github.com/apache/tvm/pull/20167

   Reduce peak WASM memory during BF16 tensor-cache loading by decoding directly in the final `float32` CPU tensor allocation. The packed BF16 input is copied into the lower half of the destination tensor and expanded backward, preventing output writes from overwriting unread input. This removes the separate FFI byte-array allocation previously required during decoding. Leave the existing decoder as a fallback when the new WASM entry point is unavailable.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "guan404ming (via GitHub)" <gi...@apache.org>.
guan404ming commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3849753104


##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {

Review Comment:
   Maybe a plain product over `shape` would be enough here. tensor shapes are already validated at allocation time, so the overflow checks feel a little heavier than needed.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "akaashrp (via GitHub)" <gi...@apache.org>.
akaashrp commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3859639705


##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {
+  size_t size = 1;
+  for (int i = 0; i < tensor->ndim; ++i) {
+    TVM_FFI_ICHECK_GE(tensor->shape[i], 0);
+    const uint64_t dim_u64 = static_cast<uint64_t>(tensor->shape[i]);
+    TVM_FFI_ICHECK_LE(dim_u64, std::numeric_limits<size_t>::max());
+    const size_t dim = static_cast<size_t>(dim_u64);
+    if (size != 0) {
+      TVM_FFI_ICHECK_LE(dim, std::numeric_limits<size_t>::max() / size);
+    }
+    size *= dim;
+  }
+  return size;
+}
+
+void CheckF32CPUTensor(const Tensor& tensor) {
+  TVM_FFI_ICHECK_EQ(tensor->device.device_type, kDLCPU);
+  TVM_FFI_ICHECK(tensor.IsContiguous());
+  TVM_FFI_ICHECK_EQ(tensor->dtype.code, kDLFloat);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.bits, 32);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.lanes, 1);
+}
+
 void ArrayDecodeStorage(Tensor cpu_arr, TVMFFIByteArray* bytes, const std::string& format,
                         const std::string& dtype) {
   TVM_FFI_ICHECK_NE(bytes, nullptr);
   const char* byte_data = bytes->data;
   const size_t byte_size = bytes->size;
   if (format == "f32-to-bf16" && dtype == "float32") {
-    const uint16_t* bf16 = reinterpret_cast<const uint16_t*>(byte_data);
-    uint32_t* data = static_cast<uint32_t*>(cpu_arr->data);
-    TVM_FFI_ICHECK(cpu_arr.IsContiguous());
-    size_t size = 1;
-    for (int i = 0; i < cpu_arr->ndim; ++i) {
-      size *= cpu_arr->shape[i];
+    CheckF32CPUTensor(cpu_arr);
+    const size_t size = GetCheckedTensorElementCount(cpu_arr);
+    TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+    TVM_FFI_ICHECK_EQ(byte_size, size * 2);
+    if (size == 0) {
+      return;
     }
-    TVM_FFI_ICHECK_EQ(size, byte_size / 2);
+    TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+    TVM_FFI_ICHECK_NE(byte_data, nullptr);
+    const uint8_t* bf16 = reinterpret_cast<const uint8_t*>(byte_data);
+    uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
     for (size_t i = 0; i < size; ++i) {
-      data[i] = static_cast<uint32_t>(bf16[i]) << 16;
+      data[4 * i] = 0;
+      data[4 * i + 1] = 0;
+      data[4 * i + 2] = bf16[2 * i];
+      data[4 * i + 3] = bf16[2 * i + 1];
     }
   } else {
     cpu_arr.CopyFromBytes(byte_data, byte_size);
   }
 }
 
+void ArrayDecodeBF16ToF32Inplace(Tensor cpu_arr, int64_t encoded_nbytes) {
+  CheckF32CPUTensor(cpu_arr);
+  TVM_FFI_ICHECK_GE(encoded_nbytes, 0);
+
+  const size_t size = GetCheckedTensorElementCount(cpu_arr);
+  TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+  TVM_FFI_ICHECK_EQ(static_cast<uint64_t>(encoded_nbytes), static_cast<uint64_t>(size) * 2);
+  if (size == 0) {
+    return;
+  }
+  TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+
+  uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
+  for (size_t i = size; i != 0; --i) {

Review Comment:
   Done. This now loads the BF16 bits with `memcpy` and writes each float as one 32-bit value.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "guan404ming (via GitHub)" <gi...@apache.org>.
guan404ming commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3849753101


##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {
+  size_t size = 1;
+  for (int i = 0; i < tensor->ndim; ++i) {
+    TVM_FFI_ICHECK_GE(tensor->shape[i], 0);
+    const uint64_t dim_u64 = static_cast<uint64_t>(tensor->shape[i]);
+    TVM_FFI_ICHECK_LE(dim_u64, std::numeric_limits<size_t>::max());
+    const size_t dim = static_cast<size_t>(dim_u64);
+    if (size != 0) {
+      TVM_FFI_ICHECK_LE(dim, std::numeric_limits<size_t>::max() / size);
+    }
+    size *= dim;
+  }
+  return size;
+}
+
+void CheckF32CPUTensor(const Tensor& tensor) {
+  TVM_FFI_ICHECK_EQ(tensor->device.device_type, kDLCPU);
+  TVM_FFI_ICHECK(tensor.IsContiguous());
+  TVM_FFI_ICHECK_EQ(tensor->dtype.code, kDLFloat);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.bits, 32);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.lanes, 1);
+}
+
 void ArrayDecodeStorage(Tensor cpu_arr, TVMFFIByteArray* bytes, const std::string& format,
                         const std::string& dtype) {
   TVM_FFI_ICHECK_NE(bytes, nullptr);
   const char* byte_data = bytes->data;
   const size_t byte_size = bytes->size;
   if (format == "f32-to-bf16" && dtype == "float32") {
-    const uint16_t* bf16 = reinterpret_cast<const uint16_t*>(byte_data);
-    uint32_t* data = static_cast<uint32_t*>(cpu_arr->data);
-    TVM_FFI_ICHECK(cpu_arr.IsContiguous());
-    size_t size = 1;
-    for (int i = 0; i < cpu_arr->ndim; ++i) {
-      size *= cpu_arr->shape[i];
+    CheckF32CPUTensor(cpu_arr);
+    const size_t size = GetCheckedTensorElementCount(cpu_arr);
+    TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+    TVM_FFI_ICHECK_EQ(byte_size, size * 2);
+    if (size == 0) {
+      return;
     }
-    TVM_FFI_ICHECK_EQ(size, byte_size / 2);
+    TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+    TVM_FFI_ICHECK_NE(byte_data, nullptr);
+    const uint8_t* bf16 = reinterpret_cast<const uint8_t*>(byte_data);
+    uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
     for (size_t i = 0; i < size; ++i) {
-      data[i] = static_cast<uint32_t>(bf16[i]) << 16;
+      data[4 * i] = 0;
+      data[4 * i + 1] = 0;
+      data[4 * i + 2] = bf16[2 * i];
+      data[4 * i + 3] = bf16[2 * i + 1];
     }
   } else {
     cpu_arr.CopyFromBytes(byte_data, byte_size);
   }
 }
 
+void ArrayDecodeBF16ToF32Inplace(Tensor cpu_arr, int64_t encoded_nbytes) {
+  CheckF32CPUTensor(cpu_arr);
+  TVM_FFI_ICHECK_GE(encoded_nbytes, 0);
+
+  const size_t size = GetCheckedTensorElementCount(cpu_arr);
+  TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+  TVM_FFI_ICHECK_EQ(static_cast<uint64_t>(encoded_nbytes), static_cast<uint64_t>(size) * 2);
+  if (size == 0) {
+    return;
+  }
+  TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+
+  uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
+  for (size_t i = size; i != 0; --i) {

Review Comment:
   I think this would be a bit faster if we wrote each element as one 32-bit value (e.g. via `memcpy` of a `uint32_t`) instead of four separate byte stores, since this loop runs over every BF16 parameter.



##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {

Review Comment:
   I think a plain product over `shape` would be enough here; tensor shapes are already validated at allocation time, so the overflow checks feel a little heavier than needed.



##########
web/src/runtime.ts:
##########
@@ -1373,6 +1417,24 @@ export class Instance implements Disposable {
     }
   }
 
+  /** Return the exact byte size of a packed BF16 tensor. */
+  private getPackedBF16Bytes(shape: Array<number>): number {

Review Comment:
   I think we could drop this helper, since `ArrayDecodeBF16ToF32Inplace` already rejects a size mismatch via `TVM_FFI_ICHECK_EQ`. Keeping the check in one place would be a bit simpler to maintain.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "akaashrp (via GitHub)" <gi...@apache.org>.
akaashrp commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3860052682


##########
web/src/runtime.ts:
##########
@@ -1373,6 +1417,24 @@ export class Instance implements Disposable {
     }
   }
 
+  /** Return the exact byte size of a packed BF16 tensor. */
+  private getPackedBF16Bytes(shape: Array<number>): number {

Review Comment:
   The C++ check only runs after `storeRawBytes` has copied the packed data into tensor memory, so it cannot protect that copy. I think the JS check is needed before the write. Passing the bytes through `ArrayDecodeStorage` would reintroduce the temporary allocation this path removes. I prefer to keep the C++ check as validation of the newly exposed FFI entry point itself since it can be invoked independently of the tensor-cache loader.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "guan404ming (via GitHub)" <gi...@apache.org>.
guan404ming commented on PR #20167:
URL: https://github.com/apache/tvm/pull/20167#issuecomment-5426688180

   Thanks for your update. Currently looks nice!


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "guan404ming (via GitHub)" <gi...@apache.org>.
guan404ming commented on PR #20167:
URL: https://github.com/apache/tvm/pull/20167#issuecomment-5426693947

   @tvm-bot rerun


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "akaashrp (via GitHub)" <gi...@apache.org>.
akaashrp merged PR #20167:
URL: https://github.com/apache/tvm/pull/20167


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "guan404ming (via GitHub)" <gi...@apache.org>.
guan404ming commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3849753101


##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {
+  size_t size = 1;
+  for (int i = 0; i < tensor->ndim; ++i) {
+    TVM_FFI_ICHECK_GE(tensor->shape[i], 0);
+    const uint64_t dim_u64 = static_cast<uint64_t>(tensor->shape[i]);
+    TVM_FFI_ICHECK_LE(dim_u64, std::numeric_limits<size_t>::max());
+    const size_t dim = static_cast<size_t>(dim_u64);
+    if (size != 0) {
+      TVM_FFI_ICHECK_LE(dim, std::numeric_limits<size_t>::max() / size);
+    }
+    size *= dim;
+  }
+  return size;
+}
+
+void CheckF32CPUTensor(const Tensor& tensor) {
+  TVM_FFI_ICHECK_EQ(tensor->device.device_type, kDLCPU);
+  TVM_FFI_ICHECK(tensor.IsContiguous());
+  TVM_FFI_ICHECK_EQ(tensor->dtype.code, kDLFloat);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.bits, 32);
+  TVM_FFI_ICHECK_EQ(tensor->dtype.lanes, 1);
+}
+
 void ArrayDecodeStorage(Tensor cpu_arr, TVMFFIByteArray* bytes, const std::string& format,
                         const std::string& dtype) {
   TVM_FFI_ICHECK_NE(bytes, nullptr);
   const char* byte_data = bytes->data;
   const size_t byte_size = bytes->size;
   if (format == "f32-to-bf16" && dtype == "float32") {
-    const uint16_t* bf16 = reinterpret_cast<const uint16_t*>(byte_data);
-    uint32_t* data = static_cast<uint32_t*>(cpu_arr->data);
-    TVM_FFI_ICHECK(cpu_arr.IsContiguous());
-    size_t size = 1;
-    for (int i = 0; i < cpu_arr->ndim; ++i) {
-      size *= cpu_arr->shape[i];
+    CheckF32CPUTensor(cpu_arr);
+    const size_t size = GetCheckedTensorElementCount(cpu_arr);
+    TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+    TVM_FFI_ICHECK_EQ(byte_size, size * 2);
+    if (size == 0) {
+      return;
     }
-    TVM_FFI_ICHECK_EQ(size, byte_size / 2);
+    TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+    TVM_FFI_ICHECK_NE(byte_data, nullptr);
+    const uint8_t* bf16 = reinterpret_cast<const uint8_t*>(byte_data);
+    uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
     for (size_t i = 0; i < size; ++i) {
-      data[i] = static_cast<uint32_t>(bf16[i]) << 16;
+      data[4 * i] = 0;
+      data[4 * i + 1] = 0;
+      data[4 * i + 2] = bf16[2 * i];
+      data[4 * i + 3] = bf16[2 * i + 1];
     }
   } else {
     cpu_arr.CopyFromBytes(byte_data, byte_size);
   }
 }
 
+void ArrayDecodeBF16ToF32Inplace(Tensor cpu_arr, int64_t encoded_nbytes) {
+  CheckF32CPUTensor(cpu_arr);
+  TVM_FFI_ICHECK_GE(encoded_nbytes, 0);
+
+  const size_t size = GetCheckedTensorElementCount(cpu_arr);
+  TVM_FFI_ICHECK_LE(size, std::numeric_limits<size_t>::max() / 4);
+  TVM_FFI_ICHECK_EQ(static_cast<uint64_t>(encoded_nbytes), static_cast<uint64_t>(size) * 2);
+  if (size == 0) {
+    return;
+  }
+  TVM_FFI_ICHECK_NE(cpu_arr->data, nullptr);
+
+  uint8_t* data = static_cast<uint8_t*>(cpu_arr->data) + cpu_arr->byte_offset;
+  for (size_t i = size; i != 0; --i) {

Review Comment:
   This would be a bit faster if we wrote each element as one 32-bit value (e.g. via `memcpy` of a `uint32_t`) instead of four separate byte stores, since this loop runs over every BF16 parameter.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "akaashrp (via GitHub)" <gi...@apache.org>.
akaashrp commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3860052682


##########
web/src/runtime.ts:
##########
@@ -1373,6 +1417,24 @@ export class Instance implements Disposable {
     }
   }
 
+  /** Return the exact byte size of a packed BF16 tensor. */
+  private getPackedBF16Bytes(shape: Array<number>): number {

Review Comment:
   The C++ check only runs after `storeRawBytes` has copied the packed data into tensor memory, so it cannot protect that copy. The JS check is needed before the write. Passing the bytes through `ArrayDecodeStorage` would reintroduce the temporary allocation this path removes. I would keep the C++ check as validation of the newly exposed FFI entry point itself since it can be invoked independently of the tensor-cache loader.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Web] Decode packed BF16 tensor records in place [tvm]

Posted by "akaashrp (via GitHub)" <gi...@apache.org>.
akaashrp commented on code in PR #20167:
URL: https://github.com/apache/tvm/pull/20167#discussion_r3860060026


##########
web/emcc/wasm_runtime.cc:
##########
@@ -126,38 +128,93 @@ TVM_FFI_STATIC_INIT_BLOCK() {
       });
 }
 
+size_t GetCheckedTensorElementCount(const Tensor& tensor) {

Review Comment:
   I couldn't find shape validation in the allocation path. `ffi::GetDataSize` multiplies dimensions as `size_t` without checking negative values or overflow. This matters on `wasm32`, where an `int64_t` dimension could truncate before allocation. Since this function is also callable directly through FFI, I think the checks should remain unless validation is added to the common tensor construction path.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org