Changing name of parameters from factual_decay to top_p_decay

megatron/text_generation/api.py

@@ -33,8 +33,8 @@ def generate_and_post_process(model,
                               return_output_log_probs=False,
                               top_k_sampling=0,
                               top_p_sampling=0.0,
-                              factual_decay=0.0,
-                              factual_bound=0.0,
+                              top_p_decay=0.0,
+                              top_p_bound=0.0,
                               temperature=1.0,
                               add_BOS=False,
                               use_eod_token_for_early_termination=True,
@@ -52,8 +52,8 @@ def generate_and_post_process(model,
         return_output_log_probs=return_output_log_probs,
         top_k_sampling=top_k_sampling,
         top_p_sampling=top_p_sampling,
-        factual_decay=factual_decay,
-        factual_bound=factual_bound,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
         temperature=temperature,
         add_BOS=add_BOS,
         use_eod_token_for_early_termination=use_eod_token_for_early_termination,
@@ -82,8 +82,8 @@ def generate(model,
              return_output_log_probs=False,
              top_k_sampling=0,
              top_p_sampling=0.0,
-             factual_decay=0.0,
-             factual_bound=0.0,
+             top_p_decay=0.0,
+             top_p_bound=0.0,
              temperature=1.0,
              add_BOS=False,
              use_eod_token_for_early_termination=True,
@@ -101,7 +101,7 @@ def generate(model,
     # Make sure input params are avaialble to all ranks.
     values = [tokens_to_generate,
               return_output_log_probs,
-              top_k_sampling, top_p_sampling, factual_decay, factual_bound,
+              top_k_sampling, top_p_sampling, top_p_decay, top_p_bound,
               temperature, add_BOS, use_eod_token_for_early_termination,
               stop_on_double_eol,
               stop_on_eol,
@@ -111,8 +111,8 @@ def generate(model,
     return_output_log_probs = bool(values_float_tensor[1].item())
     top_k_sampling = int(values_float_tensor[2].item())
     top_p_sampling = values_float_tensor[3].item()
-    factual_decay = values_float_tensor[4].item()
-    factual_bound = values_float_tensor[5].item()
+    top_p_decay = values_float_tensor[4].item()
+    top_p_bound = values_float_tensor[5].item()
     temperature = values_float_tensor[6].item()
     add_BOS = bool(values_float_tensor[7].item())
     use_eod_token_for_early_termination = bool(values_float_tensor[8].item())
@@ -142,8 +142,8 @@ def generate(model,
         return_output_log_probs=return_output_log_probs,
         top_k=top_k_sampling,
         top_p=top_p_sampling,
-        factual_decay=factual_decay,
-        factual_bound=factual_bound,
+        top_p_decay=top_p_decay,
+        top_p_bound=top_p_bound,
         temperature=temperature,
         use_eod_token_for_early_termination=use_eod_token_for_early_termination,
         stop_on_double_eol=stop_on_double_eol,

megatron/text_generation/generation.py

@@ -94,7 +94,7 @@ def score_and_return_on_first_stage(model, tokens, lengths):
 def generate_tokens_probs_and_return_on_first_stage(
         model, tokens, lengths,
         return_output_log_probs=False,
-        top_k=0, top_p=0.0, factual_decay=0.0, factual_bound=0.0,
+        top_k=0, top_p=0.0, top_p_decay=0.0, top_p_bound=0.0,
         temperature=1.0,
         use_eod_token_for_early_termination=True,
         stop_on_double_eol=False,
@@ -200,10 +200,10 @@ def generate_tokens_probs_and_return_on_first_stage(
                                     top_p=top_p,
                                     temperature=temperature,
                                     vocab_size=tokenizer.vocab_size)
-                if top_p > 0.0 and factual_decay > 0.0:
-                    top_p = top_p * factual_decay
-                    if factual_bound > 0.0:
-                        top_p = max(top_p, factual_bound)
+                if top_p > 0.0 and top_p_decay > 0.0:
+                    top_p = top_p * top_p_decay
+                    if top_p_bound > 0.0:
+                        top_p = max(top_p, top_p_bound)
 
                 # If a prompt length is smaller or equal th current context
                 # length, it means we have started generating tokens

megatron/text_generation_server.py

@@ -93,25 +93,25 @@ class MegatronGenerate(Resource):
             if not (0 <= top_p <= 1.0):
                 return "top_p must be less than or equal to 1.0"
         
-        factual_decay = 0.0
-        if "factual_decay" in request.get_json():
-            factual_decay = request.get_json()["factual_decay"]
-            if not (type(factual_decay) == float):
-                return "factual_decay must be a positive float less than or equal to 1.0"
+        top_p_decay = 0.0
+        if "top_p_decay" in request.get_json():
+            top_p_decay = request.get_json()["top_p_decay"]
+            if not (type(top_p_decay) == float):
+                return "top_p_decay must be a positive float less than or equal to 1.0"
             if top_p == 0.0:
-                return "factual_decay cannot be set without top_p"
-            if not (0 <= factual_decay <= 1.0):
-                return "factual_decay must be less than or equal to 1.0"
-        
-        factual_bound = 0.0
-        if "factual_bound" in request.get_json():
-            factual_bound = request.get_json()["factual_bound"]
-            if not (type(factual_bound) == float):
-                return "factual_bound must be a positive float less than or equal to top_p"
+                return "top_p_decay cannot be set without top_p"
+            if not (0 <= top_p_decay <= 1.0):
+                return "top_p_decay must be less than or equal to 1.0"
+        
+        top_p_bound = 0.0
+        if "top_p_bound" in request.get_json():
+            top_p_bound = request.get_json()["top_p_bound"]
+            if not (type(top_p_bound) == float):
+                return "top_p_bound must be a positive float less than or equal to top_p"
             if top_p == 0.0:
-                return "factual_bound cannot be set without top_p"
-            if not (0.0 < factual_bound <= top_p):
-                return "factual_bound must be greater than 0 and less than top_p"
+                return "top_p_bound cannot be set without top_p"
+            if not (0.0 < top_p_bound <= top_p):
+                return "top_p_bound must be greater than 0 and less than top_p"
         
         add_BOS = False
         if "add_BOS" in request.get_json():
@@ -163,8 +163,8 @@ class MegatronGenerate(Resource):
                         return_output_log_probs=logprobs,
                         top_k_sampling=top_k,
                         top_p_sampling=top_p,
-                        factual_decay=factual_decay,
-                        factual_bound=factual_bound,
+                        top_p_decay=top_p_decay,
+                        top_p_bound=top_p_bound,
                         temperature=temperature,
                         add_BOS=add_BOS,
                         use_eod_token_for_early_termination=True,