add gpu flag

neulab · Mar 18, 2022 · 0c1d86c · 0c1d86c
1 parent 8afab92
commit 0c1d86c
Show file tree

Hide file tree

Showing 2 changed files with 29 additions and 17 deletions.
diff --git a/fairseq/knnlm.py b/fairseq/knnlm.py
@@ -1,5 +1,6 @@
 import torch
 import faiss
+import faiss.contrib.torch_utils
 import math
 import numpy as np
 from fairseq import utils
@@ -27,16 +28,25 @@ def setup_faiss(self, args):
         print('Reading datastore took {} s'.format(time.time() - start))
         index.nprobe = args.probe
 
+        if args.knnlm_gpu:
+            start = time.time()
+            co = faiss.GpuClonerOptions()
+            co.useFloat16 = True
+            index = faiss.index_cpu_to_gpu(faiss.StandardGpuResources(), 0, index, co)
+            print('Moving index to GPU took {} s'.format(time.time() - start))
+
+        dstore_float_dtype = np.float32
+        dstore_int_dtype = np.int32
         if args.dstore_fp16:
-            print('Keys are fp16 and vals are int16')
-            if not args.no_load_keys:
-                self.keys = np.memmap(args.dstore_filename+'_keys.npy', dtype=np.float16, mode='r', shape=(self.dstore_size, self.dimension))
-            self.vals = np.memmap(args.dstore_filename+'_vals.npy', dtype=np.int16, mode='r', shape=(self.dstore_size, 1))
-        else:
-            print('Keys are fp32 and vals are int64')
-            if not args.no_load_keys:
-                self.keys = np.memmap(args.dstore_filename+'_keys.npy', dtype=np.float32, mode='r', shape=(self.dstore_size, self.dimension))
-            self.vals = np.memmap(args.dstore_filename+'_vals.npy', dtype=np.int, mode='r', shape=(self.dstore_size, 1))
+            print('Keys are fp16')
+            dstore_float_dtype = np.float16
+
+        if not args.no_load_keys:
+            self.keys = np.memmap(args.dstore_filename+'_keys.npy', dtype=dstore_float_dtype, mode='r',
+                                  shape=(self.dstore_size, self.dimension))
+        self.vals = np.memmap(args.dstore_filename+'_vals.npy', dtype=np.int64, mode='r',
+                              shape=(self.dstore_size, 1))
+        # self.vals = torch.from_numpy(self.vals).to(self.device)
 
         # If you wish to load all the keys into memory
         # CAUTION: Only do this if your RAM can handle it!
@@ -46,16 +56,17 @@ def setup_faiss(self, args):
 
             if not args.no_load_keys:
                 del self.keys
-                self.keys_from_memmap = np.memmap(args.dstore_filename+'_keys.npy', dtype=np.float32, mode='r', shape=(self.dstore_size, self.dimension))
-                self.keys = np.zeros((self.dstore_size, self.dimension), dtype=np.float16 if args.dstore_fp16 else np.float32)
+                self.keys_from_memmap = np.memmap(args.dstore_filename+'_keys.npy', dtype=dstore_float_dtype, mode='r', shape=(self.dstore_size, self.dimension))
+                self.keys = np.zeros((self.dstore_size, self.dimension), dtype=dstore_float_dtype)
                 self.keys = self.keys_from_memmap[:]
-                self.keys = self.keys.astype(np.float16 if args.dstore_fp16 else np.float32)
+                self.keys = self.keys.astype(dstore_float_dtype)
 
             del self.vals
-            self.vals_from_memmap = np.memmap(args.dstore_filename+'_vals.npy', dtype=np.int, mode='r', shape=(self.dstore_size, 1))
-            self.vals = np.zeros((self.dstore_size, 1), dtype=np.int16 if args.dstore_fp16 else np.int)
-            self.vals = self.vals_from_memmap[:]
-            self.vals = self.vals.astype(np.int16 if args.dstore_fp16 else np.int)
+            vals_from_memmap = np.memmap(args.dstore_filename+'_vals.npy', dtype=np.int, mode='r', shape=(self.dstore_size, 1))
+            self.vals = np.zeros((self.dstore_size, 1), dtype=np.int64)
+            self.vals = vals_from_memmap[:]
+            self.vals = self.vals.astype(dstore_int_dtype)
+            del vals_from_memmap
             print('Loading to memory took {} s'.format(time.time() - start))
 
         return index
@@ -110,7 +121,7 @@ def dist_func(d, k, q, function=None):
 
         # (T_reducedxB)
         yhat_knn_prob = torch.logsumexp(probs + index_mask, dim=-1).clone()
-        full_yhat_knn_prob = torch.full([qshape[0]*qshape[1]], -10000).cuda()
+        full_yhat_knn_prob = torch.full([qshape[0]*qshape[1]], -10000.0).cuda()
         full_yhat_knn_prob[tgt != pad_idx] = yhat_knn_prob
 
         # TxBx1

diff --git a/fairseq/options.py b/fairseq/options.py
@@ -487,6 +487,7 @@ def add_eval_lm_args(parser):
                        help='helpful for certain ops that are only used during eval')
     group.add_argument('--knnlm', action='store_true',
                        help='use the k-nearest neighbors language model')
+    group.add_argument('--knnlm-gpu', action='store_true')
     group.add_argument('--save-knnlm-dstore', action='store_true',
                        help='save keys for the knnlm datastore')
     group.add_argument('--dstore-mmap', default=None, type=str,