Fix the forwarding of the length penalty parameter (IBM#70)

It was erroneously tied to the presence of the repetition penalty
parameter. So when the repetition penalty wasn't set, the server would
get this parameter even though the client sent it to the router.

---------

Signed-off-by: Max de Bayser <[email protected]>

Author: maxdebayser

integration_tests/test_cases_bloom560m.yaml

@@ -1254,7 +1254,7 @@
 
 
 # Length penalty
-- name: Length penalty
+- name: Length penalty with repetition penalty
   request:
     params:
       decoding:
@@ -1273,6 +1273,24 @@
         stopReason: EOS_TOKEN
         text: The first time I saw the movie, it was in
 
+# Length penalty
+- name: Length penalty
+  request:
+    params:
+      decoding:
+        length_penalty:
+          start_index: 8
+          decay_factor: 1.01
+      stopping:
+        maxNewTokens: 20
+    requests:
+      - {"text": "A very long story:\n"}
+  response:
+    responses:
+      - generatedTokenCount: 12
+        inputTokenCount: 6
+        stopReason: EOS_TOKEN
+        text: The first time I saw the movie, I was a
 
 # Multiple inputs
 - name: Multiple inputs

integration_tests/test_cases_mt0small.yaml

@@ -1166,7 +1166,7 @@
 
 
 # Length penalty
-- name: Length penalty
+- name: Length penalty with repetition penalty
   request:
     params:
       decoding:
@@ -1186,6 +1186,24 @@
         text: The very long story is
 
 
+# Length penalty
+- name: Length penalty
+  request:
+    params:
+      decoding:
+        length_penalty:
+          start_index: 3
+          decay_factor: 4.0
+      stopping:
+        maxNewTokens: 20
+    requests:
+      - {"text": "A very long story:\n"}
+  response:
+    responses:
+      - generatedTokenCount: 7
+        inputTokenCount: 8
+        stopReason: EOS_TOKEN
+        text: The very long story is
 
 # Multiple inputs
 - name: Multiple inputs

router/src/grpc_server.rs

@@ -545,9 +545,9 @@ fn convert_params(
             if let Some(d) = p.decoding {
                 if d.repetition_penalty != 0.0 {
                     gp.repetition_penalty = d.repetition_penalty;
-                    gp.length_penalty =
-                        d.length_penalty.map(|lp| (lp.start_index, lp.decay_factor));
                 }
+                gp.length_penalty =
+                    d.length_penalty.map(|lp| (lp.start_index, lp.decay_factor));
             }
             // Stopping Criteria
             if let Some(s) = p.stopping {