Add MCP control server, live API-generation view, and repetition penalty
- scripts/jwash_mcp.py: MCP server (FastMCP/stdio), an HTTP client of the running J-Wash server so an external LLM can drive an already-loaded model. Tools: generate, find_token, list_layers, scale_token/replace_token (pure-weights, layers required), set_intensity, list_edits/reset_edits. - api/app.py + ui: /api/generate now records the last exchange (surfaced in /api/status) and broadcasts on /ws when done; the UI shows a "generated via API/MCP" panel at the top of the chat and an Options "API monitor" toggle that swaps the 2s status poll for an event-driven refresh. - sampling: repetition penalty (default 1.0, applied in model_manager._sample), exposed as a "rep" field in the chat controls; intentionally not exposed through the MCP. - ui: remove the redundant "md" chat toggle (already available in Options). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
407ce9eef0
commit
4bc7c8005e
@@ -0,0 +1,299 @@
|
||||
"""MCP server for J-Wash — let an external LLM autonomously test token-direction
|
||||
edits on a model that is ALREADY loaded in the running J-Wash app (port 8381).
|
||||
|
||||
It exposes only what is needed to experiment, and nothing else:
|
||||
|
||||
* generate — (re)generate text from the current model
|
||||
* scale_token / replace_token — apply a pure-weights token operation, any intensity
|
||||
* set_intensity — global multiplier over all edits (sweep the intensity)
|
||||
* list_edits / reset_edits — inspect / clear the current edits
|
||||
|
||||
This is a thin HTTP client of the J-Wash REST API (same server as scripts/jlab.py),
|
||||
spoken over MCP/stdio so any MCP client (Claude Desktop, another agent, ...) can
|
||||
drive a model you loaded yourself. By design it never loads models or lenses,
|
||||
never changes the sampling defaults beyond the call, and never exports anything:
|
||||
a model AND a Jacobian lens must already be loaded from the J-Wash UI.
|
||||
|
||||
Token edits are always applied in a *pure-weights* mode (read projection, or W_U
|
||||
abliteration on Gemma-style models): the live preview matches an exported
|
||||
checkpoint exactly, so what the model tests here is what a baked model would do.
|
||||
|
||||
Run it from an MCP client over stdio:
|
||||
|
||||
pip install mcp
|
||||
python -X utf8 scripts/jwash_mcp.py
|
||||
|
||||
Point it at a non-default J-Wash instance with an env var:
|
||||
|
||||
JWASH_BASE=http://127.0.0.1:8382
|
||||
"""
|
||||
import json
|
||||
import os
|
||||
import urllib.error
|
||||
import urllib.parse
|
||||
import urllib.request
|
||||
|
||||
from mcp.server.fastmcp import FastMCP
|
||||
|
||||
BASE = os.environ.get("JWASH_BASE", "http://127.0.0.1:8381").rstrip("/")
|
||||
|
||||
mcp = FastMCP(
|
||||
"j-wash",
|
||||
instructions=(
|
||||
"Drive a model ALREADY loaded in the running J-Wash app to test "
|
||||
"token-direction edits. Typical loop: (1) `generate` a baseline reply; "
|
||||
"(2) find the exact token with `find_token` and the layers to target with "
|
||||
"`list_layers`; (3) apply edits with `scale_token`/`replace_token` (layers "
|
||||
"are required) — always pure-weights, faithful to an exported checkpoint; "
|
||||
"(4) `generate` again to see the effect, tuning each edit's `factor` or the "
|
||||
"global `set_intensity`; (5) `reset_edits` to start over. A model AND a "
|
||||
"Jacobian lens must be loaded from the J-Wash UI first; this server never "
|
||||
"loads models, lenses, or exports checkpoints."
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
# --- HTTP plumbing (stdlib only, like scripts/jlab.py) ----------------------
|
||||
|
||||
def _call(method, path, body=None, timeout=600):
|
||||
url = BASE + path
|
||||
data = json.dumps(body).encode("utf-8") if body is not None else None
|
||||
req = urllib.request.Request(url, data=data, method=method)
|
||||
if data is not None:
|
||||
req.add_header("Content-Type", "application/json")
|
||||
try:
|
||||
with urllib.request.urlopen(req, timeout=timeout) as resp:
|
||||
return json.loads(resp.read().decode("utf-8"))
|
||||
except urllib.error.HTTPError as exc:
|
||||
detail = exc.read().decode("utf-8", "replace")
|
||||
try:
|
||||
detail = json.loads(detail).get("detail", detail)
|
||||
except (json.JSONDecodeError, AttributeError):
|
||||
pass
|
||||
raise ValueError(f"J-Wash {method} {path} -> HTTP {exc.code}: {detail}")
|
||||
except urllib.error.URLError as exc:
|
||||
raise ValueError(
|
||||
f"J-Wash server unreachable at {BASE} ({exc.reason}). Start it "
|
||||
"(python -X utf8 run.py) and load a model + lens, then retry."
|
||||
)
|
||||
|
||||
|
||||
def _status():
|
||||
return _call("GET", "/api/status")
|
||||
|
||||
|
||||
def _resolve_token(text):
|
||||
"""The EXACT single token for ``text`` (a leading space is significant)."""
|
||||
r = _call("GET", "/api/token-lookup?q=" + urllib.parse.quote(text.strip()))
|
||||
cands = r.get("candidates", [])
|
||||
for c in cands:
|
||||
if c["str"] == text:
|
||||
return c
|
||||
listing = ", ".join(f"{c['id']}:{c['str']!r}" for c in cands) or "none"
|
||||
raise ValueError(
|
||||
f"No exact single-token match for {text!r}. A leading space is "
|
||||
f"significant (mid-sentence words usually need one, e.g. ' model'). "
|
||||
f"Candidates: {listing}"
|
||||
)
|
||||
|
||||
|
||||
def _parse_layers(spec, n_layers):
|
||||
"""None -> server default band; 'all'/'none'/'19-31'/'3,5,7' -> explicit list."""
|
||||
if spec is None:
|
||||
return None
|
||||
spec = spec.strip().lower()
|
||||
if spec in ("", "none"):
|
||||
return []
|
||||
if spec == "all":
|
||||
if not n_layers:
|
||||
raise ValueError("layers='all' needs a loaded model to know the layer count")
|
||||
return list(range(n_layers))
|
||||
out = set()
|
||||
for part in spec.split(","):
|
||||
part = part.strip()
|
||||
if not part:
|
||||
continue
|
||||
if "-" in part:
|
||||
lo, hi = part.split("-", 1)
|
||||
out.update(range(int(lo), int(hi) + 1))
|
||||
else:
|
||||
out.add(int(part))
|
||||
return sorted(out)
|
||||
|
||||
|
||||
_PURE_WEIGHTS_MODES = ("readthrough", "exact", "abliteration")
|
||||
|
||||
|
||||
def _edits_summary():
|
||||
st = _status()
|
||||
mode = st.get("interventions_mode")
|
||||
return {
|
||||
"mode": mode,
|
||||
"pure_weights": mode in _PURE_WEIGHTS_MODES,
|
||||
"global_intensity": st.get("interventions_scale"),
|
||||
"edits": [
|
||||
{
|
||||
"id": r["id"],
|
||||
"token": r["token"],
|
||||
"op": r["mode"],
|
||||
"factor": r["factor"],
|
||||
"replacement": r.get("replacement"),
|
||||
"layers": r["layers"],
|
||||
}
|
||||
for r in (st.get("interventions") or [])
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def _add_rule(token, op, factor, replacement, layers):
|
||||
st = _status()
|
||||
loaded = st.get("loaded")
|
||||
if not loaded:
|
||||
raise ValueError(
|
||||
"No model loaded in J-Wash — load a model and a Jacobian lens from "
|
||||
"the app first."
|
||||
)
|
||||
if not st.get("lens"):
|
||||
raise ValueError(
|
||||
"No Jacobian lens loaded — load one in the Lens tab of J-Wash before "
|
||||
"editing tokens."
|
||||
)
|
||||
# Force a pure-weights mode: read projection, or W_U abliteration on
|
||||
# architectures that normalize their writes (Gemma 2/3 style).
|
||||
pure_mode = "abliteration" if loaded.get("rebase_supported") is False else "readthrough"
|
||||
_call("PATCH", "/api/interventions", {"mode": pure_mode})
|
||||
|
||||
body = {"token_id": _resolve_token(token)["id"], "mode": op, "factor": float(factor)}
|
||||
if op == "replace":
|
||||
body["replacement_id"] = _resolve_token(replacement)["id"]
|
||||
parsed = _parse_layers(layers, loaded.get("n_layers"))
|
||||
if parsed is not None:
|
||||
body["layers"] = parsed
|
||||
_call("POST", "/api/interventions", body)
|
||||
return _edits_summary()
|
||||
|
||||
|
||||
# --- MCP tools --------------------------------------------------------------
|
||||
|
||||
@mcp.tool()
|
||||
def generate(prompt: str, system: str | None = None, max_tokens: int = 200,
|
||||
temperature: float = 0.0, seed: int = 1234) -> str:
|
||||
"""(Re)generate a reply from the model currently loaded in J-Wash, with the
|
||||
active token edits applied — call it again to regenerate.
|
||||
|
||||
At temperature 0 generation is deterministic, so the reply changes only when
|
||||
the edits change: this is the clean way to compare behaviour before vs after
|
||||
an edit. Raise `temperature` (or set `seed=-1` for a random seed) to sample
|
||||
varied continuations instead. `system` is an optional system prompt.
|
||||
"""
|
||||
messages = [{"role": "system", "content": system}] if system else []
|
||||
messages.append({"role": "user", "content": prompt})
|
||||
r = _call("POST", "/api/generate", {
|
||||
"messages": messages,
|
||||
"sampling": {"temperature": temperature, "max_tokens": max_tokens, "seed": seed},
|
||||
}, timeout=1800)
|
||||
return r.get("text", "")
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def scale_token(token: str, factor: float, layers: str) -> dict:
|
||||
"""Multiply a token's own direction by `factor` (pure-weights edit).
|
||||
|
||||
`factor` is the intensity: 0 removes the token's direction, 0<factor<1
|
||||
attenuates it, factor>1 amplifies it. `token` is the exact token string — a
|
||||
leading space is usually significant (e.g. ' model'); use `find_token` to get
|
||||
it. `layers` is REQUIRED: it selects where the edit acts and an edit that
|
||||
targets no layer does nothing — pass a 0-based range or list ('19-25', '20',
|
||||
'20,24', or 'all') and call `list_layers` to see the model's layers. The mode
|
||||
is forced to pure-weights so the effect matches an exported checkpoint.
|
||||
"""
|
||||
return _add_rule(token, "scale", factor, None, layers)
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def replace_token(token: str, replacement: str, layers: str, factor: float = 1.0) -> dict:
|
||||
"""Rewrite `token`'s component onto `replacement`'s direction (pure-weights),
|
||||
e.g. token=' model', replacement=' fish' to make the model talk like a fish.
|
||||
|
||||
You MUST pass `layers` — it selects the layers where the replacement is
|
||||
applied, and WITHOUT it nothing happens. Give a 0-based range or list
|
||||
('19-25', '20,24', or 'all'); call `list_layers` for the model's layers and
|
||||
`find_token` for the exact ' token' strings (both must be single tokens, a
|
||||
leading space usually being significant). `factor` scales the strength
|
||||
(1.0 = full). The mode is forced to pure-weights (faithful to an export).
|
||||
"""
|
||||
return _add_rule(token, "replace", factor, replacement, layers)
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def set_intensity(scale: float) -> dict:
|
||||
"""Set the global multiplier applied to ALL active edits — sweep the overall
|
||||
intensity without touching each rule: 0 disables every edit, 1 is nominal,
|
||||
>1 pushes them harder. Returns the current edits.
|
||||
"""
|
||||
_call("PATCH", "/api/interventions", {"scale": scale})
|
||||
return _edits_summary()
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def list_edits() -> dict:
|
||||
"""Show the active token edits, the pure-weights mode in force, and the global
|
||||
intensity — a read-only snapshot of the current experiment.
|
||||
"""
|
||||
return _edits_summary()
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def reset_edits() -> dict:
|
||||
"""Remove every token edit, returning the model to its unedited behaviour.
|
||||
Use it to start a fresh experiment. Returns the (now empty) edits.
|
||||
"""
|
||||
_call("DELETE", "/api/interventions")
|
||||
return _edits_summary()
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def find_token(text: str) -> dict:
|
||||
"""Look up the single-token forms of `text` so you can pick the exact token to
|
||||
edit before calling scale_token/replace_token.
|
||||
|
||||
A leading space is significant (' model' and 'model' are different tokens), so
|
||||
the lookup also tries the space-prefixed and capitalization variants and
|
||||
returns those that are exactly one token. Use a returned `token` string
|
||||
verbatim; words that split into several tokens can't be edited directly.
|
||||
"""
|
||||
r = _call("GET", "/api/token-lookup?q=" + urllib.parse.quote(text.strip()))
|
||||
return {
|
||||
"query": text,
|
||||
"candidates": [{"id": c["id"], "token": c["str"]} for c in r.get("candidates", [])],
|
||||
}
|
||||
|
||||
|
||||
@mcp.tool()
|
||||
def list_layers() -> dict:
|
||||
"""List the layers you can target with `scale_token`/`replace_token`.
|
||||
|
||||
Returns the model's total layer count (indices are 0-based, so the valid
|
||||
range is 0..n_layers-1) and the layers the loaded Jacobian lens actually
|
||||
covers — those are the calibrated ones to edit; targeting a layer outside
|
||||
them falls back to a less reliable logit-lens direction.
|
||||
"""
|
||||
st = _status()
|
||||
loaded = st.get("loaded")
|
||||
if not loaded:
|
||||
raise ValueError(
|
||||
"No model loaded in J-Wash — load a model and a Jacobian lens from "
|
||||
"the app first."
|
||||
)
|
||||
n = loaded.get("n_layers")
|
||||
lens = st.get("lens") or {}
|
||||
out = {"n_layers": n, "valid_range": f"0-{n - 1}" if n else None}
|
||||
if lens.get("fitted_layers_all"):
|
||||
out["lens_fitted_layers"] = lens["fitted_layers_all"]
|
||||
if lens.get("tapped_layers"):
|
||||
out["lens_tapped_layers"] = lens["tapped_layers"]
|
||||
return out
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
mcp.run(transport="stdio")
|
||||
Reference in New Issue
Block a user