| 1 |
"""Dialect → Go `text/template` registry. |
| 2 |
|
| 3 |
One `.gotmpl` file per chat-template dialect, shipped beside this |
| 4 |
module. The registry also records per-dialect: |
| 5 |
|
| 6 |
- **Default stops** — token strings that Ollama must refuse to |
| 7 |
generate past. Missing stops cause runaway generation (findings §9). |
| 8 |
- **Default params** — `temperature` + `top_p` tuned per family. |
| 9 |
|
| 10 |
The templates themselves are Go `text/template` syntax: `{{.Role}}`, |
| 11 |
`{{range .Messages}}`, `{{if .System}}`, etc. Ollama renders them |
| 12 |
at inference time. The round-trip test (`tests/unit/export/ollama/ |
| 13 |
test_template_registry_roundtrip.py`) verifies each template |
| 14 |
produces token-identical output to the base model's Jinja reference |
| 15 |
on a fixed message-set matrix. |
| 16 |
""" |
| 17 |
|
| 18 |
from __future__ import annotations |
| 19 |
|
| 20 |
from dataclasses import dataclass, field |
| 21 |
from pathlib import Path |
| 22 |
from typing import Final, Literal |
| 23 |
|
| 24 |
from dlm.export.ollama.errors import TemplateRegistryError |
| 25 |
|
| 26 |
Dialect = Literal[ |
| 27 |
"chatml", |
| 28 |
"qwen3thinking", |
| 29 |
"gemma2", |
| 30 |
"smollm3", |
| 31 |
"olmo2", |
| 32 |
"llama3", |
| 33 |
"phi3", |
| 34 |
"phi4mini", |
| 35 |
"mistral", |
| 36 |
] |
| 37 |
|
| 38 |
_TEMPLATES_DIR: Final[Path] = Path(__file__).resolve().parent / "templates" |
| 39 |
|
| 40 |
|
| 41 |
@dataclass(frozen=True) |
| 42 |
class DialectTemplate: |
| 43 |
"""One registry row: a dialect's Go template + param defaults.""" |
| 44 |
|
| 45 |
dialect: Dialect |
| 46 |
template_path: Path |
| 47 |
default_stops: tuple[str, ...] |
| 48 |
default_temperature: float = 0.7 |
| 49 |
default_top_p: float = 0.9 |
| 50 |
# Added-by-training special tokens go here ONLY if they're |
| 51 |
# dialect-inherent (e.g., chatml's `<|im_end|>`). Per-adapter added |
| 52 |
# tokens from pad fallback come from the adapter |
| 53 |
# tokenizer at render time. |
| 54 |
extra_stop_hints: tuple[str, ...] = field(default_factory=tuple) |
| 55 |
|
| 56 |
def read_template(self) -> str: |
| 57 |
"""Return the `.gotmpl` file contents verbatim.""" |
| 58 |
if not self.template_path.is_file(): |
| 59 |
raise TemplateRegistryError( |
| 60 |
f"template file missing for {self.dialect!r}: {self.template_path}" |
| 61 |
) |
| 62 |
return self.template_path.read_text(encoding="utf-8") |
| 63 |
|
| 64 |
|
| 65 |
_REGISTRY: Final[dict[Dialect, DialectTemplate]] = { |
| 66 |
"chatml": DialectTemplate( |
| 67 |
dialect="chatml", |
| 68 |
template_path=_TEMPLATES_DIR / "chatml.gotmpl", |
| 69 |
# Qwen 2.5 variants emit role-delimiter tokens like |
| 70 |
# `<|im_start|>` at the top of each turn. Listing them as stops |
| 71 |
# prevents runaway prompt-continuation when the model tries to |
| 72 |
# synthesize a new turn instead of yielding. |
| 73 |
default_stops=("<|im_end|>", "<|endoftext|>", "<|im_start|>"), |
| 74 |
), |
| 75 |
"qwen3thinking": DialectTemplate( |
| 76 |
dialect="qwen3thinking", |
| 77 |
template_path=_TEMPLATES_DIR / "qwen3thinking.gotmpl", |
| 78 |
# Qwen3's reasoning profile still uses ChatML turn framing, but |
| 79 |
# the upstream defaults run slightly broader sampling than the |
| 80 |
# legacy ChatML family. |
| 81 |
default_stops=("<|im_end|>", "<|endoftext|>", "<|im_start|>"), |
| 82 |
default_temperature=0.6, |
| 83 |
default_top_p=0.95, |
| 84 |
), |
| 85 |
"gemma2": DialectTemplate( |
| 86 |
dialect="gemma2", |
| 87 |
template_path=_TEMPLATES_DIR / "gemma2.gotmpl", |
| 88 |
# Gemma 2 instruct uses `<start_of_turn>` / `<end_of_turn>` |
| 89 |
# framing and a trailing `<start_of_turn>model` generation |
| 90 |
# prompt. Stop on both turn-boundary tokens and `<eos>` so |
| 91 |
# Ollama yields instead of continuing into a fresh role block. |
| 92 |
default_stops=("<end_of_turn>", "<eos>", "<start_of_turn>"), |
| 93 |
), |
| 94 |
"smollm3": DialectTemplate( |
| 95 |
dialect="smollm3", |
| 96 |
template_path=_TEMPLATES_DIR / "smollm3.gotmpl", |
| 97 |
# SmolLM3 keeps ChatML turn framing but ships a reasoning-first |
| 98 |
# instruct prompt. Stop on the turn delimiters so Ollama |
| 99 |
# yields instead of hallucinating another role block. |
| 100 |
default_stops=("<|im_end|>", "<|end_of_text|>", "<|im_start|>"), |
| 101 |
default_temperature=0.6, |
| 102 |
default_top_p=0.95, |
| 103 |
), |
| 104 |
"olmo2": DialectTemplate( |
| 105 |
dialect="olmo2", |
| 106 |
template_path=_TEMPLATES_DIR / "olmo2.gotmpl", |
| 107 |
# OLMo2 uses `<|user|>` / `<|assistant|>` role markers and |
| 108 |
# closes assistant turns with `<|endoftext|>`. |
| 109 |
default_stops=("<|endoftext|>", "<|user|>", "<|assistant|>", "<|system|>"), |
| 110 |
), |
| 111 |
"llama3": DialectTemplate( |
| 112 |
dialect="llama3", |
| 113 |
template_path=_TEMPLATES_DIR / "llama3.gotmpl", |
| 114 |
# Llama-3 instruct uses `<|start_header_id|>` to begin a role |
| 115 |
# block; include it so the model can't generate a spurious new |
| 116 |
# turn. `<|eom_id|>` is Llama-3.1's tool-call terminator. |
| 117 |
default_stops=( |
| 118 |
"<|eot_id|>", |
| 119 |
"<|end_of_text|>", |
| 120 |
"<|start_header_id|>", |
| 121 |
"<|eom_id|>", |
| 122 |
), |
| 123 |
), |
| 124 |
"phi3": DialectTemplate( |
| 125 |
dialect="phi3", |
| 126 |
template_path=_TEMPLATES_DIR / "phi3.gotmpl", |
| 127 |
# Phi-3.5-mini's chat template uses `<|user|>`, `<|assistant|>`, |
| 128 |
# `<|system|>` role delimiters — the audit called these out. |
| 129 |
default_stops=( |
| 130 |
"<|end|>", |
| 131 |
"<|endoftext|>", |
| 132 |
"<|user|>", |
| 133 |
"<|assistant|>", |
| 134 |
"<|system|>", |
| 135 |
), |
| 136 |
), |
| 137 |
"phi4mini": DialectTemplate( |
| 138 |
dialect="phi4mini", |
| 139 |
template_path=_TEMPLATES_DIR / "phi4mini.gotmpl", |
| 140 |
# Phi-4-mini-reasoning keeps Phi-style `<|role|>` markers and |
| 141 |
# `<|end|>` closers, but the upstream template always injects a |
| 142 |
# default system preamble before user turns. |
| 143 |
default_stops=( |
| 144 |
"<|end|>", |
| 145 |
"<|endoftext|>", |
| 146 |
"<|user|>", |
| 147 |
"<|assistant|>", |
| 148 |
"<|system|>", |
| 149 |
), |
| 150 |
), |
| 151 |
"mistral": DialectTemplate( |
| 152 |
dialect="mistral", |
| 153 |
template_path=_TEMPLATES_DIR / "mistral.gotmpl", |
| 154 |
# Mistral's instruct format wraps user turns in `[INST] ... [/INST]`; |
| 155 |
# `[INST]` as a stop prevents the model from restarting a turn. |
| 156 |
default_stops=("</s>", "[/INST]", "[INST]"), |
| 157 |
), |
| 158 |
} |
| 159 |
|
| 160 |
|
| 161 |
def registered_dialects() -> tuple[Dialect, ...]: |
| 162 |
"""Tuple of all shipped dialects. Useful for CLI help + parametrize.""" |
| 163 |
return tuple(_REGISTRY.keys()) |
| 164 |
|
| 165 |
|
| 166 |
def get_template(dialect: str) -> DialectTemplate: |
| 167 |
"""Return the registry row for `dialect` or raise. |
| 168 |
|
| 169 |
`dialect` is typed as `str` at the edge so callers from pydantic |
| 170 |
`Literal` fields can pass through; we validate against the |
| 171 |
registry keys. |
| 172 |
""" |
| 173 |
if dialect not in _REGISTRY: |
| 174 |
raise TemplateRegistryError( |
| 175 |
f"unknown template dialect {dialect!r}; registered: {sorted(_REGISTRY)}" |
| 176 |
) |
| 177 |
return _REGISTRY[dialect] |
| 178 |
|
| 179 |
|
| 180 |
def load_template_text(dialect: str) -> str: |
| 181 |
"""Shortcut: `get_template(dialect).read_template()`.""" |
| 182 |
return get_template(dialect).read_template() |