minor bug fix

eth-easl · Oct 13, 2024 · 87fc48e · 87fc48e
1 parent fece40b
commit 87fc48e
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 8 deletions.
diff --git a/tests/ops/test_sbmm.py b/tests/ops/test_sbmm.py
@@ -61,11 +61,13 @@ def run_problem(
         except torch.cuda.OutOfMemoryError as e:
             print(f"Out of memory, skipping nr={nr}, nm={nm}, m={m}, k={k}")
         finally:
-            torch.cuda.empty_cache()
+            pass
+            # torch.cuda.empty_cache()
 
     def test_tiny(self):
-        for i in range(10000):
-            self.run_problem("uniform",  10,  5, 256,  256)
+        for i in range(20):
+            self.run_problem("uniform",  10,  5, 256,  512)
+            # self.run_problem("uniform", 128, 2, 4096, 12288)
             self.run_problem("zipf:1.5", 128, 2, 4096, 12288)
 
     # def test_llama(self):

diff --git a/triteia/csrc/ops/triteia_nm_sbmm.cu b/triteia/csrc/ops/triteia_nm_sbmm.cu
@@ -748,7 +748,7 @@ __global__ void SBMM_2_4(
         meta + weight_indices * (prob_n / 16) * (prob_k / 8);
 
     const int4 *__restrict__ s_ptr = s + weight_indices * (prob_n / 8);
-    int4 *__restrict__ C_ptr = C + start * (prob_n / 8);
+    int4 *C_ptr = C + start * (prob_n / 8);
     int *locks_ptr = locks + batch_id * (prob_n / 8);
     int thread_m = -1;
     int thread_k = -1;
@@ -769,6 +769,7 @@ __global__ void SBMM_2_4(
 
     for (int i = 0; i < tot_n_blocks; i += 4) {
       int thread_n_blocks = tot_n_blocks - i;
+      count = tot_n - 16 * i;
       int par = 1;
       if (thread_n_blocks > 4) {
         par = (16 * thread_n_blocks - pad) / 64;
@@ -798,10 +799,10 @@ __global__ void SBMM_2_4(
       }
       cudaError_t err = cudaGetLastError();
       if (err != cudaSuccess) printf("Error: %s\n", cudaGetErrorString(err));
-      __syncthreads();
       A_ptr += 16 * thread_n_blocks * (prob_k / 8) * par;
       C_ptr += 16 * thread_n_blocks * (prob_n / 8) * par;
     }
+    C_ptr = C + start * (prob_n / 8);
   }
 };
 

diff --git a/triteia/python/ops/matmul/sbmm.py b/triteia/python/ops/matmul/sbmm.py
@@ -93,10 +93,8 @@ def sbmm_4bit_2_4_native(qweights, xs, metas, ss, indices, base_weight=None):
         y = torch.zeros(xs.shape[0], ss.shape[2], dtype=xs.dtype, device=xs.device)
     if torch.all(indices == -1):
         return y
-
     unique_indices, counts = torch.unique_consecutive(indices, return_counts=True)
     if len(unique_indices) == 1:
-        # use a normal matmul
         workspace = torch.zeros(
             y.shape[1] // 128 * 16, device=xs.device, dtype=torch.int32
         )
@@ -127,7 +125,7 @@ def sbmm_4bit_2_4_native(qweights, xs, metas, ss, indices, base_weight=None):
             len(unique_indices), y.shape[1] // 8, device=xs.device, dtype=torch.int32
         )
         output = torch.zeros(
-            (xs.shape[0], y.shape[1]), dtype=torch.float16, device=xs.device
+            (xs.shape[0], ss.shape[2]), dtype=torch.float16, device=xs.device
         )
         sbmm_2_4(
             xs,