mirror of
https://github.com/cloudstack-llc/mlx-knife.git
synced 2026-07-21 18:15:37 -04:00
Nikesh Parajuli
58 lines
2.7 KiB
Diff
58 lines
2.7 KiB
Diff
diff --git a/mlxk2/core/runner/__init__.py b/mlxk2/core/runner/__init__.py
|
|
index d22aacc..ca9ad06 100644
|
|
--- a/mlxk2/core/runner/__init__.py
|
|
+++ b/mlxk2/core/runner/__init__.py
|
|
@@ -553,7 +553,7 @@ class MLXRunner:
|
|
# Use sliding window for proper decoding
|
|
start_idx = max(0, len(generated_tokens) - context_window)
|
|
window_tokens = generated_tokens[start_idx:]
|
|
- window_text = self._decode_tokens(window_tokens)
|
|
+ window_text = self.tokenizer.decode(window_tokens)
|
|
|
|
# Extract new text
|
|
if start_idx == 0:
|
|
@@ -565,13 +565,13 @@ class MLXRunner:
|
|
new_text = window_text
|
|
previous_decoded = window_text
|
|
else:
|
|
- new_text = self._decode_tokens(window_tokens)
|
|
+ new_text = self.tokenizer.decode(window_tokens)
|
|
if len(window_tokens) > 1:
|
|
- prefix = self._decode_tokens(window_tokens[:-1])
|
|
+ prefix = self.tokenizer.decode(window_tokens[:-1])
|
|
if new_text.startswith(prefix):
|
|
new_text = new_text[len(prefix):]
|
|
else:
|
|
- new_text = self._decode_tokens([token_id])
|
|
+ new_text = self.tokenizer.decode([token_id])
|
|
|
|
if new_text:
|
|
accumulated_response += new_text
|
|
@@ -750,7 +750,7 @@ class MLXRunner:
|
|
# Decode full response using the streaming detokenizer
|
|
# This properly converts BPE space markers (Ġ U+0120) to spaces (U+0020)
|
|
# while tokenizer.decode() for slow tokenizers (LlamaTokenizer) does NOT.
|
|
- full_response = self._decode_tokens(all_tokens)
|
|
+ full_response = self.tokenizer.decode(all_tokens)
|
|
|
|
# Debug: Show raw generated tokens for quality analysis (enabled via --verbose)
|
|
if self.verbose:
|
|
@@ -762,7 +762,7 @@ class MLXRunner:
|
|
for tid in last_3_ids:
|
|
try:
|
|
# Use detokenizer for debug output too
|
|
- decoded = self._decode_tokens([tid])
|
|
+ decoded = self.tokenizer.decode([tid])
|
|
last_3_decoded.append(f"{tid}={decoded!r}")
|
|
except Exception:
|
|
last_3_decoded.append(f"{tid}=<error>")
|
|
@@ -778,7 +778,7 @@ class MLXRunner:
|
|
response = full_response[len(formatted_prompt):]
|
|
else:
|
|
# Decode generated tokens only (use detokenizer)
|
|
- decoded = self._decode_tokens(generated_tokens)
|
|
+ decoded = self.tokenizer.decode(generated_tokens)
|
|
response = decoded if isinstance(decoded, str) else str(decoded)
|
|
|
|
# Filter stop tokens (strings only)
|