Manually pin protobuf version and remove two quantization args

unavailable in Praxis 1.0.0.

requirements-cuda.txt

@@ -508,7 +508,7 @@ prompt-toolkit==3.0.39
     # via
     #   ipython
     #   jupyter-console
-protobuf==3.19.6
+protobuf==3.20.3
     # via
     #   googleapis-common-protos
     #   grpcio-reflection

requirements.txt

@@ -506,7 +506,7 @@ prompt-toolkit==3.0.39
     # via
     #   ipython
     #   jupyter-console
-protobuf==3.19.6
+protobuf==3.20.3
     # via
     #   googleapis-common-protos
     #   grpcio-reflection

saxml/server/pax/quantization.py

@@ -56,8 +56,6 @@ def for_transformer(
     quantize_ngrammer_embedding: bool = False,
     dtype: jnp.dtype = jnp.int8,
     block_size: int = 0,
-    use_int4_packed_weights: bool = True,
-    int4_packed_weights_container_dtype: jnp.dtype = jnp.int32,
 ):
   """Find and quantize transformer.
 
@@ -86,11 +84,6 @@ def for_transformer(
       Ngrammer/VQNgrammer layer.
     dtype: Dtype of the quantized variables.
     block_size: Block size for sub-channel quantization. Defaults to off.
-    use_int4_packed_weights: If True, pack/unpack int4 weights into int32 or
-      int8. It is for int4 weights only and has not effect on other type. If
-      False int4 weights will be kept in int8.
-    int4_packed_weights_container_dtype: Container type for int4 weights: int32
-      to pack 8 int4s, or int8 to pack 2 int4s.
 
   Returns:
     a modifier that quantizes transformers when applied to a config.
@@ -130,8 +123,6 @@ def task(self):
             quantize_ngrammer_embedding=quantize_ngrammer_embedding,
             dtype=dtype,
             block_size=block_size,
-            use_int4_packed_weights=use_int4_packed_weights,
-            int4_packed_weights_container_dtype=int4_packed_weights_container_dtype,
         )
         return task_p