Add blocksize=64 4-bit quantization support for ROCm CDNA (warp64) GPUs

bitsandbytes/backends/cuda/ops.py

@@ -212,7 +212,7 @@ def _(A: torch.Tensor, code: torch.Tensor, blocksize: int) -> tuple[torch.Tensor
     torch._check_is_size(blocksize)
 
     if ROCM_WARP_SIZE_64:
-        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128])
+        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64])
     else:
         torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64, 32])
 
@@ -270,7 +270,7 @@ def _dequantize_blockwise_impl(
     A: torch.Tensor, absmax: torch.Tensor, code: torch.Tensor, blocksize: int, dtype: torch.dtype, out: torch.Tensor
 ) -> None:
     if ROCM_WARP_SIZE_64:
-        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128])
+        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64])
     else:
         torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64, 32])
 
@@ -304,7 +304,7 @@ def _(
     A: torch.Tensor, blocksize: int, quant_type: str, quant_storage: torch.dtype
 ) -> tuple[torch.Tensor, torch.Tensor]:
     if ROCM_WARP_SIZE_64:
-        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128])
+        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64])
     else:
         torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64, 32])
 
@@ -386,7 +386,7 @@ def _dequantize_4bit_impl(
     out: torch.Tensor,
 ) -> None:
     if ROCM_WARP_SIZE_64:
-        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128])
+        torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64])
     else:
         torch._check(blocksize in [4096, 2048, 1024, 512, 256, 128, 64, 32])
 

csrc/kernels.hip

@@ -443,6 +443,94 @@ __global__ void kQuantizeBlockwise(float * code, T * __restrict__ const A, float
   }
 }
 
+// Specialized kernel for blocksize=64 with 4-bit quantization
+// Works on both warp32 and warp64 hardware
+// Processes 2 blocks of 64 values per thread block using 64 threads
+// Uses logical warps of 32: threads 0-31 handle block 0, threads 32-63 handle block 1
+//   - warp32: 2 hardware warps, each reduces naturally
+//   - warp64: 1 hardware warp split into 2 logical warps of 32
+template <typename T, int DATA_TYPE>
+__global__ void kQuantizeBlockwise64(
+    float* code, T* __restrict__ const A, float* absmax, unsigned char* out, float* __restrict__ const rand,
+    const int rand_offset, const int n
+) {
+    constexpr int BLOCK_SIZE = 64;        // Size of each quantization block
+    constexpr int NUM_PER_TH = 2;         // Values per thread (for 4-bit packing)
+    constexpr int THREADS = 64;           // Total threads per HIP block
+    constexpr int THREADS_PER_BLOCK = 32; // Threads handling each quantization block
+
+    const int base_idx = blockIdx.x * BLOCK_SIZE * 2; // 2 quantization blocks per HIP block
+
+    T vals[NUM_PER_TH];
+    unsigned char qvals[NUM_PER_TH / 2]; // For 4-bit: 2 values per byte
+    float local_abs_max = 0.0f;
+
+    const int block_id = threadIdx.x / THREADS_PER_BLOCK;        // 0 for threads 0-31, 1 for threads 32-63
+    const int local_thread_id = threadIdx.x % THREADS_PER_BLOCK; // Thread ID within the quantization block (0-31)
+
+    typedef hipcub::BlockLoad<T, THREADS, NUM_PER_TH, hipcub::BLOCK_LOAD_WARP_TRANSPOSE> LoadT;
+    typedef hipcub::BlockStore<unsigned char, THREADS, NUM_PER_TH / 2, hipcub::BLOCK_STORE_WARP_TRANSPOSE> StoreChar;
+    // Logical warp size of 32: on warp32 this matches hardware warps,
+    // on warp64 this splits the single hardware warp into two independent reductions
+    typedef hipcub::WarpReduce<float, 32> WarpReduce;
+
+    __shared__ typename LoadT::TempStorage loadt;
+    __shared__ typename StoreChar::TempStorage storec;
+    __shared__ typename WarpReduce::TempStorage warp_reduce[2]; // One per logical warp
+    __shared__ float smem_absmax_value[2];
+
+    const int i = base_idx + block_id * BLOCK_SIZE;
+    // Use a flag instead of early return: BlockLoad/BlockStore/__syncthreads are cooperative
+    // operations that require ALL 64 threads to participate
+    const bool block_valid = (i < n);
+
+    // All 64 threads participate in the load (out-of-bounds threads get 0.0f)
+    __syncthreads();
+    LoadT(loadt).Load(&(A[base_idx]), vals, min(BLOCK_SIZE * 2, n - base_idx), (T)0.0f);
+
+    // Each thread computes max of its values
+    local_abs_max = -FLT_MAX;
+#pragma unroll NUM_PER_TH
+    for (int j = 0; j < NUM_PER_TH; j++)
+        local_abs_max = fmaxf(local_abs_max, fabsf((float)vals[j]));
+
+    // Reduce within each logical warp of 32 threads independently
+    local_abs_max = WarpReduce(warp_reduce[block_id]).Reduce(local_abs_max, hipcub::Max());
+
+    if (local_thread_id == 0) {
+        if (block_valid) {
+            smem_absmax_value[block_id] = 1.0f / local_abs_max;
+            absmax[blockIdx.x * 2 + block_id] = local_abs_max;
+        } else {
+            smem_absmax_value[block_id] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    local_abs_max = smem_absmax_value[block_id];
+
+    switch (DATA_TYPE) {
+    case FP4:
+#pragma unroll NUM_PER_TH
+        for (int j = 0; j < NUM_PER_TH / 2; j++) {
+            qvals[j] = dQuantizeFP4(((float)vals[2 * j]) * local_abs_max) << 4;
+            qvals[j] |= dQuantizeFP4(((float)vals[2 * j + 1]) * local_abs_max);
+        }
+        break;
+    case NF4:
+#pragma unroll NUM_PER_TH
+        for (int j = 0; j < NUM_PER_TH / 2; j++) {
+            qvals[j] = dQuantizeNF4(((float)vals[2 * j]) * local_abs_max) << 4;
+            qvals[j] |= dQuantizeNF4(((float)vals[2 * j + 1]) * local_abs_max);
+        }
+        break;
+    }
+
+    // All 64 threads participate in the store (valid_items limits the actual writes)
+    __syncthreads();
+    StoreChar(storec).Store(&(out[base_idx / 2]), qvals, min((BLOCK_SIZE * 2 + 1) / 2, (n - base_idx + 1) / 2));
+}
+
 template<typename T, int TILE_SIZE, int THREADS, int NUM_PER_TH, int DATA_TYPE>
 __global__ void kDequantizeBlockwise(float *code, unsigned char * A, float * absmax, T *out, const int blocksize, const int n)
 {
@@ -2566,6 +2654,20 @@ MAKE_kQuantizeBlockwise(hip_bfloat16,  128, 2, 0, NF4)
   MAKE_kQuantizeBlockwise(hip_bfloat16,   64, 2, 0, NF4)
 #endif
 
+// Specialized blocksize=64 4-bit quantization kernel instantiations for ROCm
+#define MAKE_kQuantizeBlockwise64(dtype, data_type_name) \
+template __global__ void kQuantizeBlockwise64<dtype, data_type_name>(float * code, dtype * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+
+// FP4 instantiations
+MAKE_kQuantizeBlockwise64(half, FP4)
+MAKE_kQuantizeBlockwise64(float, FP4)
+MAKE_kQuantizeBlockwise64(hip_bfloat16, FP4)
+
+// NF4 instantiations
+MAKE_kQuantizeBlockwise64(half, NF4)
+MAKE_kQuantizeBlockwise64(float, NF4)
+MAKE_kQuantizeBlockwise64(hip_bfloat16, NF4)
+
 template __global__ void kDequantizeBlockwise<half, 512, 64, 8, FP4>(float *code, unsigned char * A, float * absmax, half *out, const int blocksize, const int n);
 template __global__ void kDequantizeBlockwise<half, 512, 64, 8, General8bit>(float *code, unsigned char * A, float * absmax, half *out, const int blocksize, const int n);
 template __global__ void kDequantizeBlockwise<half, 512, 64, 8, NF4>(float *code, unsigned char * A, float * absmax, half *out, const int blocksize, const int n);

csrc/kernels_hip.cuh

@@ -19,6 +19,11 @@ __global__ void kQuantizeBlockwise(
     float* code, T* __restrict__ const A, float* absmax, unsigned char* out, float* __restrict__ const rand,
     const int rand_offset, const int n
 );
+template <typename T, int DATA_TYPE>
+__global__ void kQuantizeBlockwise64(
+    float* code, T* __restrict__ const A, float* absmax, unsigned char* out, float* __restrict__ const rand,
+    const int rand_offset, const int n
+);
 template <typename T, int BLOCK_SIZE, int THREADS, int NUM_PER_TH, int DATA_TYPE>
 __global__ void
     kDequantizeBlockwise(float* code, unsigned char* A, float* absmax, T* out, const int blocksize, const int n);

csrc/ops.hip

@@ -56,9 +56,14 @@ template <typename T, int STOCHASTIC, int DATA_TYPE> void quantizeBlockwise(floa
    hipLaunchKernelGGL(( kQuantizeBlockwise<T, 256, 2, 0, DATA_TYPE>), dim3(num_blocks), dim3(128), 0, 0, code, A, absmax, out, rand, rand_offset, n);
   else if(blocksize == 128)
    hipLaunchKernelGGL(( kQuantizeBlockwise<T, 128, 2, 0, DATA_TYPE>), dim3(num_blocks), dim3(64), 0, 0, code, A, absmax, out, rand, rand_offset, n);
-  else if(blocksize == 64 && BNB_WARP_SIZE == 32)
-   hipLaunchKernelGGL(( kQuantizeBlockwise<T, 64, 2, 0, DATA_TYPE>), dim3(num_blocks), dim3(32), 0, 0, code, A, absmax, out, rand, rand_offset, n);
-
+  else if(blocksize == 64) {
+    // For 4-bit (FP4/NF4): use specialized kernel that processes 2 blocks of 64 per thread block
+    // Works on all warp sizes (32 and 64) by using logical warps of 32
+    if constexpr(DATA_TYPE > 0)
+      hipLaunchKernelGGL(( kQuantizeBlockwise64<T, DATA_TYPE>), dim3((num_blocks + 1) / 2), dim3(64), 0, 0, code, A, absmax, out, rand, rand_offset, n);
+    else
+      hipLaunchKernelGGL(( kQuantizeBlockwise<T, 64, 2, 0, DATA_TYPE>), dim3(num_blocks), dim3(32), 0, 0, code, A, absmax, out, rand, rand_offset, n);
+  }
 
   CUDA_CHECK_RETURN(hipPeekAtLastError());
 }

tests/test_functional.py

@@ -1098,7 +1098,7 @@ class TestQuantize4BitFunctional:
     @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
     @pytest.mark.parametrize(
         "blocksize",
-        [32, 64, 128, 256, 512, 1024, 2048, 4096] if not ROCM_WARP_SIZE_64 else [128, 256, 512, 1024, 2048, 4096],
+        [32, 64, 128, 256, 512, 1024, 2048, 4096] if not ROCM_WARP_SIZE_64 else [64, 128, 256, 512, 1024, 2048, 4096],
     )
     def test_4bit_quant(self, device, dtype, quant_type, blocksize):
         if device == "hpu" and not is_supported_on_hpu(quant_type, dtype):
@@ -1173,7 +1173,7 @@ def test_4bit_quant(self, device, dtype, quant_type, blocksize):
     @pytest.mark.parametrize("device", get_available_devices())
     @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
     @pytest.mark.parametrize(
-        "blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [128], ids=id_formatter("blocksize")
+        "blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [64, 128], ids=id_formatter("blocksize")
     )
     @pytest.mark.parametrize("dtype", [torch.float32, torch.float16], ids=describe_dtype)
     def test_4bit_compressed_stats(self, device, quant_type, blocksize, dtype):
@@ -1212,7 +1212,7 @@ def test_4bit_compressed_stats(self, device, quant_type, blocksize, dtype):
     @pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16], ids=describe_dtype)
     @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
     @pytest.mark.parametrize(
-        "blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [128], ids=id_formatter("blocksize")
+        "blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [64, 128], ids=id_formatter("blocksize")
     )
     def test_4bit_quant_large(self, device, dtype, quant_type, blocksize):
         """

tests/test_linear4bit.py

@@ -193,7 +193,7 @@ def test_linear_serialization(
 
 @pytest.mark.parametrize("device", get_available_devices())
 @pytest.mark.parametrize("quant_type", ["nf4", "fp4"])
-@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [128])
+@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [64, 128])
 @pytest.mark.parametrize("compress_statistics", TRUE_FALSE, ids=id_formatter("compress_statistics"))
 def test_copy_param(device, quant_type, blocksize, compress_statistics):
     if device == "hpu" and not is_supported_on_hpu(quant_type):
@@ -250,7 +250,7 @@ def test_params4bit_torch_chunk_split(device, quant_type):
 
 @pytest.mark.parametrize("device", get_available_devices())
 @pytest.mark.parametrize("quant_type", ["nf4", "fp4"])
-@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [128])
+@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [64, 128])
 @pytest.mark.parametrize("compress_statistics", TRUE_FALSE, ids=id_formatter("compress_statistics"))
 def test_deepcopy_param(device, quant_type, blocksize, compress_statistics):
     if device == "hpu" and not is_supported_on_hpu(quant_type):
@@ -279,7 +279,7 @@ def test_deepcopy_param(device, quant_type, blocksize, compress_statistics):
 
 @pytest.mark.parametrize("device", get_available_devices())
 @pytest.mark.parametrize("quant_type", ["nf4", "fp4"])
-@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [128])
+@pytest.mark.parametrize("blocksize", [32, 64, 128] if not ROCM_WARP_SIZE_64 else [64, 128])
 @pytest.mark.parametrize("compress_statistics", TRUE_FALSE, ids=id_formatter("compress_statistics"))
 def test_params4bit_real_serialization(device, quant_type, blocksize, compress_statistics):
     if device == "hpu" and not is_supported_on_hpu(quant_type):

tests/test_ops.py

@@ -152,7 +152,7 @@ class Test4bitBlockwiseQuantOps:
     @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=id_formatter("dtype"))
     @pytest.mark.parametrize("storage_dtype", [torch.uint8, torch.bfloat16], ids=id_formatter("storage_dtype"))
     @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
-    @pytest.mark.parametrize("blocksize", [32, 64, 128, 256, 512] if not ROCM_WARP_SIZE_64 else [128, 256, 512])
+    @pytest.mark.parametrize("blocksize", [32, 64, 128, 256, 512] if not ROCM_WARP_SIZE_64 else [64, 128, 256, 512])
     def test_quantize_4bit(self, device, dtype, storage_dtype, quant_type, blocksize):
         if device == "hpu" and not is_supported_on_hpu(quant_type, dtype, storage_dtype):
             pytest.skip("This configuration is not supported on HPU.")
@@ -176,7 +176,7 @@ def test_quantize_4bit(self, device, dtype, storage_dtype, quant_type, blocksize
     @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32], ids=id_formatter("dtype"))
     @pytest.mark.parametrize("storage_dtype", [torch.uint8, torch.bfloat16], ids=id_formatter("storage_dtype"))
     @pytest.mark.parametrize("quant_type", ["fp4", "nf4"])
-    @pytest.mark.parametrize("blocksize", [32, 64, 128, 256, 512] if not ROCM_WARP_SIZE_64 else [128, 256, 512])
+    @pytest.mark.parametrize("blocksize", [32, 64, 128, 256, 512] if not ROCM_WARP_SIZE_64 else [64, 128, 256, 512])
     def test_dequantize_4bit(self, device, dtype, storage_dtype, quant_type, blocksize):
         if device == "hpu" and not is_supported_on_hpu(quant_type, dtype, storage_dtype):
             pytest.skip("This configuration is not supported on HPU.")