Allow dtype for bitsandbytes (it works, checked for idefics

9b/llama/80b)t

integration-tests/conftest.py

@@ -210,6 +210,7 @@ def local_launcher(
         quantize: Optional[str] = None,
         trust_remote_code: bool = False,
         use_flash_attention: bool = True,
+        dtype: Optional[str] = None
     ):
         port = random.randint(8000, 10_000)
         master_port = random.randint(10_000, 20_000)
@@ -237,6 +238,9 @@ def local_launcher(
         if quantize is not None:
             args.append("--quantize")
             args.append(quantize)
+        if dtype is not None:
+            args.append("--dtype")
+            args.append(dtype)
         if trust_remote_code:
             args.append("--trust-remote-code")
 

integration-tests/models/test_idefics.py

@@ -3,7 +3,7 @@
 
 @pytest.fixture(scope="module")
 def idefics_handle(launcher):
-    with launcher("HuggingFaceM4/idefics-9b-instruct", num_shard=2) as handle:
+    with launcher("HuggingFaceM4/idefics-9b-instruct", num_shard=2, dtype="float16") as handle:
         yield handle
 
 

server/text_generation_server/cli.py

@@ -76,7 +76,7 @@ def serve(
     # Downgrade enum into str for easier management later on
     quantize = None if quantize is None else quantize.value
     dtype = None if dtype is None else dtype.value
-    if dtype is not None and quantize is not None:
+    if dtype is not None and quantize not in {None, "bitsandbytes", "bitsandbytes-nf4", "bitsandbytes-fp4"}:
         raise RuntimeError(
             "Only 1 can be set between `dtype` and `quantize`, as they both decide how goes the final model."
         )