feat(tito): add DeepSeek V3.2 speculative session verification

.pre-commit-config.yaml

@@ -1,6 +1,10 @@
 default_language_version:
   python: python3
 
+# The vendored upstream module stays byte-diffable against its source; keep
+# formatters and linters away so upstream-tracking diffs remain minimal.
+exclude: '^miles/utils/chat_template_utils/templates/encoding_dsv32\.py$'
+
 ci:
   autofix_prs: true
   autoupdate_commit_msg: '[pre-commit.ci] pre-commit suggestions'

miles/backends/sglang_utils/arguments.py

@@ -107,9 +107,15 @@ def new_add_argument_wrapper(*name_or_flags, **kwargs):
             # Avoid double prefixing if dest somehow already starts with sglang_
             if not original_dest.startswith("sglang_"):
                 final_kwargs["dest"] = f"sglang_{original_dest}"
-        # If 'dest' is not explicitly provided (or is None/not a string),
-        # argparse will derive 'dest' from the (now prefixed) flag names.
-        # E.g., if the first flag is "--sglang-foo-bar", argparse sets dest to "sglang_foo_bar".
+        elif "dest" not in final_kwargs:
+            # argparse derives dest from the first alias, so store parallel sizes under SGLang's short field names.
+            for item_flag in name_or_flags:
+                if not isinstance(item_flag, str) or not item_flag.startswith("--"):
+                    continue
+                canonical_dest = item_flag[2:].replace("-", "_")
+                if canonical_dest in ("tp_size", "dp_size", "pp_size", "ep_size"):
+                    final_kwargs["dest"] = f"sglang_{canonical_dest}"
+                    break
 
         old_add_argument(*new_name_or_flags_list, **final_kwargs)
 

miles/utils/chat_template_utils/deepseek.py

@@ -1,11 +1,11 @@
 """Shared bridge for the DeepSeek official-encoder families (V3.2, V4).
 
-Neither family ships a jinja chat_template: sglang renders their prompts
-through per-family ``encoding_dsv*`` modules that share one calling
-convention, and miles' ``apply_chat_template`` routes any matching tokenizer
-here so training-side renders stay byte-aligned with what the runtime
-serves.  Each family is one ``DeepSeekFamily`` instance wrapping its encoder
-module; everything else is shared.
+Neither family ships a jinja chat_template: V4 renders through sglang's
+``encoding_dsv4``, while V3.2 uses miles' vendored
+``templates.encoding_dsv32``.  Both modules share one calling convention,
+and miles' ``apply_chat_template`` routes any matching tokenizer here.  Each
+family is one ``DeepSeekFamily`` instance wrapping its encoder module;
+everything else is shared.
 """
 
 from __future__ import annotations
@@ -16,9 +16,11 @@
 import os
 from typing import Any
 
-from sglang.srt.entrypoints.openai import encoding_dsv4, encoding_dsv32
+from sglang.srt.entrypoints.openai import encoding_dsv4
 from sglang.srt.entrypoints.openai.protocol import Tool
 
+from miles.utils.chat_template_utils.templates import encoding_dsv32
+
 _ASSISTANT_SP_TOKEN = "<｜Assistant｜>"
 
 

miles/utils/chat_template_utils/templates/__init__.py

@@ -0,0 +1 @@
+"""Bundled fixed chat templates and the vendored official DeepSeek V3.2 encoder."""