checkpoint

.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+competition/.venv-competition/

Justfile

@@ -0,0 +1,2 @@
+vendor:
+    pip download -r competition/requirements-vendor.txt -d competition/wheelhouse/ --python-version 3.12 --platform manylinux --no-deps

competition/build.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+# One-time intake step: install the vendored wheels offline.
+# The sandbox has no network, so everything must come from wheelhouse/.
+set -euo pipefail
+cd "$(dirname "$0")"
+python -m pip install --no-index --find-links=wheelhouse -r requirements-vendor.txt
+echo "[build] vendored wheels installed" >&2

competition/competition.py

@@ -0,0 +1,336 @@
+"""Competition stdio agent: direct policy sampling over a trained equinox checkpoint.
+
+Self-contained competition bot. It speaks the wire protocol in
+`generals-bots/competition/protocol.py` (handshake, then one observation frame
+per turn, one action line per turn, EOF on stdin = game over) and, on each
+turn, samples an action from the policy network exactly as the model is used
+during training/evaluation — no tree search.
+
+Unlike `scripts/mcts_agent.py`, this module does **not** import the
+`general_bots_training` training package. It inlines the policy-value network
+and the observation encoding so the script can be dropped into the competition
+sandbox with just the engine (`generals`), `equinox`, `jax`, and `numpy`
+available.
+
+This is the search-free counterpart to `scripts/competition_puct.py`: it shares
+the same network and observation encoding, but selects an action by sampling
+from the masked policy (as in `PolicyValueNetwork.__call__`) instead of running
+particle PUCT.
+
+Run it directly through the bundled matchup driver, e.g.:
+
+    PYTHONPATH=src:generals-bots .venv/bin/python \
+      generals-bots/competition/matchup.py \
+      scripts/competition.py --checkpoint ppo_model.eqx \
+      generals-bots/competition/agents/expander_python/run.sh \
+      --mode competition
+
+or wrap it in a `run.sh` that passes the checkpoint path.
+"""
+
+import argparse
+import os
+import sys
+import time
+
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+import jax.random as jrandom
+import numpy as np
+from generals.core import game
+from generals.core.action import compute_valid_move_mask
+from generals.core.game import GameState
+from generals.core.observation import Observation
+from generals.modifiers import build_castles, deathtouch
+
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+
+PASS_ACTION = np.array([1, 0, 0, 0, 0], dtype=np.int32)
+
+
+# ---------------------------------------------------------------------------
+# Policy-value network (inlined from general_bots_training.network)
+# ---------------------------------------------------------------------------
+
+
+def obs_to_tensor(obs: Observation) -> jnp.ndarray:
+    """Encode an Observation into a (C, H, W) float32 tensor for the network.
+
+    Armies and army-counts are log-normalized; land counts are divided by the
+    number of cells; scalar values are broadcast to spatial planes so a plain
+    conv stack can consume them. Grid-size agnostic.
+    """
+    H, W = obs.armies.shape
+    armies = jnp.log1p(obs.armies.astype(jnp.float32)) / jnp.log(50.0)
+
+    def bcast(scalar):
+        return jnp.broadcast_to(scalar.astype(jnp.float32), (H, W))
+
+    own_land = bcast(obs.owned_land_count / (H * W))
+    own_army = bcast(jnp.log1p(obs.owned_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    opp_land = bcast(obs.opponent_land_count / (H * W))
+    opp_army = bcast(jnp.log1p(obs.opponent_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    timestep = bcast(jnp.log1p(obs.timestep.astype(jnp.float32)) / jnp.log(1201.0))
+
+    return jnp.stack(
+        [
+            armies,
+            obs.generals.astype(jnp.float32),
+            obs.castles.astype(jnp.float32),
+            obs.mountains.astype(jnp.float32),
+            obs.neutral_cells.astype(jnp.float32),
+            obs.owned_cells.astype(jnp.float32),
+            obs.opponent_cells.astype(jnp.float32),
+            obs.fog_cells.astype(jnp.float32),
+            obs.structures_in_fog.astype(jnp.float32),
+            own_land,
+            own_army,
+            opp_land,
+            opp_army,
+            timestep,
+        ],
+        axis=0,
+    )
+
+
+class PolicyValueNetwork(eqx.Module):
+    """Conv policy-value network.
+
+    Action layout: 4 full-move directions and 4 half-move (split) directions
+    per source cell, plus one global pass action. Invalid moves are masked to
+    -1e9; pass is always available.
+
+    The value head uses global average pooling over the spatial grid, so the
+    network works for any board size without reshaping linear layers.
+    """
+
+    conv1: eqx.nn.Conv2d
+    conv2: eqx.nn.Conv2d
+    conv3: eqx.nn.Conv2d
+    conv4: eqx.nn.Conv2d
+    policy_conv: eqx.nn.Conv2d
+    value_conv: eqx.nn.Conv2d
+    value_linear1: eqx.nn.Linear
+    value_linear2: eqx.nn.Linear
+
+    def __init__(self, key, in_channels: int = 14, channels=(32, 32, 32, 16)):
+        keys = jrandom.split(key, 8)
+        self.conv1 = eqx.nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1, key=keys[0])
+        self.conv2 = eqx.nn.Conv2d(channels[0], channels[1], kernel_size=3, padding=1, key=keys[1])
+        self.conv3 = eqx.nn.Conv2d(channels[1], channels[2], kernel_size=3, padding=1, key=keys[2])
+        self.conv4 = eqx.nn.Conv2d(channels[2], channels[3], kernel_size=3, padding=1, key=keys[3])
+        # 9 = 4 dirs (full) + 4 dirs (half) + 1 pass
+        self.policy_conv = eqx.nn.Conv2d(channels[3], 9, kernel_size=1, key=keys[4])
+        self.value_conv = eqx.nn.Conv2d(channels[3], 4, kernel_size=1, key=keys[5])
+        self.value_linear1 = eqx.nn.Linear(4, 64, key=keys[6])
+        self.value_linear2 = eqx.nn.Linear(64, 1, key=keys[7])
+
+    def _features(self, obs):
+        x = jax.nn.relu(self.conv1(obs))
+        x = jax.nn.relu(self.conv2(x))
+        x = jax.nn.relu(self.conv3(x))
+        x = jax.nn.relu(self.conv4(x))
+        return x
+
+    def _value_from_features(self, feat):
+        v = jax.nn.relu(self.value_conv(feat))  # (4, H, W)
+        v = v.mean(axis=(1, 2))  # (4,) global average pool
+        v = jax.nn.relu(self.value_linear1(v))
+        return self.value_linear2(v)[0]
+
+    def policy_value(self, obs, mask):
+        """Return deterministic masked policy logits and the critic value."""
+        features = self._features(obs)
+        return self._policy_logits(features, mask), self._value_from_features(features)
+
+    def _policy_logits(self, feat, mask):
+        """Return 8*H*W move logits plus one global pass logit."""
+        logits = self.policy_conv(feat)  # (9, H, W)
+        mask_t = jnp.transpose(mask, (2, 0, 1))  # (4, H, W)
+        penalty = (1.0 - mask_t) * -1e9
+        move_penalty = jnp.concatenate([penalty, penalty], axis=0)
+        move_logits = (logits[:8] + move_penalty).reshape(-1)
+        pass_logit = jnp.mean(logits[8])[None]
+        return jnp.concatenate([move_logits, pass_logit])
+
+
+# ---------------------------------------------------------------------------
+# Action selection (matches training/evaluation rollout)
+# ---------------------------------------------------------------------------
+
+
+@jax.jit
+def competition_move_mask(obs: Observation) -> jnp.ndarray:
+    """Legal move mask, identical to the one used during training/evaluation."""
+    return compute_valid_move_mask(obs.armies, obs.owned_cells, obs.mountains)
+
+
+def _decode_policy_index(index: int, height: int, width: int) -> np.ndarray:
+    cells = height * width
+    if index == 8 * cells:
+        return PASS_ACTION.copy()
+    encoded_direction, position = divmod(index, cells)
+    row, col = divmod(position, width)
+    split = int(encoded_direction >= 4)
+    direction = encoded_direction - 4 if split else encoded_direction
+    return np.array([0, row, col, direction, split], dtype=np.int32)
+
+
+def observation_from_wire(
+    timestep: int,
+    owned_land_count: int,
+    owned_army_count: int,
+    opponent_land_count: int,
+    opponent_army_count: int,
+    type_grid,
+    owner_grid,
+    army_grid,
+) -> Observation:
+    """Convert a competition wire frame into the network's Observation type."""
+    types = jnp.asarray(type_grid, dtype=jnp.int32)
+    owners = jnp.asarray(owner_grid, dtype=jnp.int32)
+    armies = jnp.asarray(army_grid, dtype=jnp.int32)
+    visible = (types != 0) & (types != 5)
+    return Observation(
+        armies=armies,
+        generals=types == 4,
+        castles=types == 3,
+        mountains=types == 2,
+        neutral_cells=visible & (owners == 0) & (types != 2),
+        owned_cells=owners == 1,
+        opponent_cells=owners == 2,
+        fog_cells=types == 0,
+        structures_in_fog=types == 5,
+        owned_land_count=jnp.int32(owned_land_count),
+        owned_army_count=jnp.int32(owned_army_count),
+        opponent_land_count=jnp.int32(opponent_land_count),
+        opponent_army_count=jnp.int32(opponent_army_count),
+        timestep=jnp.int32(timestep),
+    )
+
+
+@jax.jit
+def competition_step(state: GameState, actions: jnp.ndarray):
+    state, actions = build_castles.apply_build_actions(state, actions)
+    return deathtouch.step(state, actions, turn=800)
+
+
+class PolicyAgent:
+    """Direct policy agent that samples actions as in training/evaluation.
+
+    On each turn it computes the masked policy logits over all legal moves and
+    samples one action with `jrandom.categorical`, exactly mirroring
+    `PolicyValueNetwork.__call__` used during rollout collection.
+    """
+
+    def __init__(
+        self,
+        network: PolicyValueNetwork,
+        seed: int = 0,
+    ):
+        self.network = network
+        self.key = jrandom.PRNGKey(seed)
+        self._infer = eqx.filter_jit(
+            lambda network, tensor, mask: network.policy_value(tensor, mask)
+        )
+
+    def warmup(self, height: int, width: int) -> None:
+        """Compile shape-dependent network and competition transition kernels."""
+        grid = jnp.zeros((height, width), dtype=jnp.int32)
+        grid = grid.at[0, 0].set(1).at[height - 1, width - 1].set(2)
+        state = game.create_initial_state(grid)
+        state = state._replace(
+            armies=state.armies.at[0, 0].set(100).at[height - 1, width - 1].set(100),
+            time=jnp.int32(100),
+        )
+        obs = game.get_observation(state, 0)
+        self.policy_value(obs)
+        actions = jnp.stack([jnp.asarray(PASS_ACTION), jnp.asarray(PASS_ACTION)])
+        warmed_state, _ = competition_step(state, actions)
+        jax.block_until_ready(warmed_state)
+
+    def policy_value(self, obs: Observation):
+        mask = competition_move_mask(obs)
+        logits, value = self._infer(self.network, obs_to_tensor(obs), mask)
+        return np.asarray(logits), float(value), mask
+
+    def select(self, obs: Observation) -> tuple[np.ndarray, dict[str, float]]:
+        """Sample one action from the masked policy and report statistics."""
+        started_at = time.perf_counter()
+        logits, value, _ = self.policy_value(obs)
+        self.key, sample_key = jrandom.split(self.key)
+        index = int(jrandom.categorical(sample_key, jnp.asarray(logits)))
+        height, width = obs.armies.shape
+        action = _decode_policy_index(index, height, width)
+        elapsed_ms = (time.perf_counter() - started_at) * 1000
+        return action, {
+            "value": value,
+            "elapsed_ms": elapsed_ms,
+        }
+
+
+# ---------------------------------------------------------------------------
+# Stdio driver
+# ---------------------------------------------------------------------------
+
+
+def _read_grid(stream, height: int):
+    return [[int(value) for value in stream.readline().split()] for _ in range(height)]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("checkpoint", help="Equinox policy checkpoint")
+    parser.add_argument("--seed", type=int, default=0)
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    handshake = sys.stdin.readline()
+    if not handshake:
+        return
+    player_index, height, width = (int(value) for value in handshake.split())
+
+    network = PolicyValueNetwork(jrandom.PRNGKey(args.seed))
+    network = eqx.tree_deserialise_leaves(args.checkpoint, network)
+    agent = PolicyAgent(network, seed=args.seed)
+    agent.warmup(height, width)
+    print(f"[competition] warmup complete for {height}x{width}", file=sys.stderr, flush=True)
+
+    while True:
+        scalar_line = sys.stdin.readline()
+        if not scalar_line:
+            return
+        timestep, own_land, own_army, opponent_land, opponent_army = (
+            int(value) for value in scalar_line.split()
+        )
+        type_grid = _read_grid(sys.stdin, height)
+        owner_grid = _read_grid(sys.stdin, height)
+        army_grid = _read_grid(sys.stdin, height)
+        observation = observation_from_wire(
+            timestep,
+            own_land,
+            own_army,
+            opponent_land,
+            opponent_army,
+            type_grid,
+            owner_grid,
+            army_grid,
+        )
+
+        started_at = time.perf_counter()
+        action, stats = agent.select(observation)
+        elapsed_ms = (time.perf_counter() - started_at) * 1000
+        print(
+            f"[competition] turn={timestep} value={stats['value']:+.3f} "
+            f"elapsed_ms={elapsed_ms:.1f}",
+            file=sys.stderr,
+            flush=True,
+        )
+        print(" ".join(str(int(value)) for value in action), flush=True)
+
+
+if __name__ == "__main__":
+    main()

competition/requirements-vendor.txt

@@ -0,0 +1,3 @@
+equinox==0.13.8
+jaxtyping==0.3.11
+wadler-lindig==0.1.7

competition/run.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+set -euo pipefail
+# Competition entrypoint. The sandbox runs `run.sh` from the submission root;
+# cd to our own directory so relative paths work regardless of the cwd.
+# `python` is the sandbox's CPython 3.12 with jax/numpy pre-installed; equinox
+# and its deps are installed by build.sh from wheelhouse/.
+cd "$(dirname "$0")"
+exec python -u competition.py model.eqx

pyproject.toml

@@ -5,12 +5,20 @@ requires-python = ">=3.13"
 dependencies = [
     "equinox>=0.13.8",
     "generals-bots",
-    "jax[cuda]>=0.11.0",
+    "jax==0.11.0",
     "jaxtyping>=0.3.11",
     "omegaconf>=2.3.1",
     "optax>=0.2.8",
 ]
 
+[project.optional-dependencies]
+cuda = [
+    "jax[cuda]==0.11.0"
+]
+cuda13 = [
+    "jax[cuda13]==0.11.0"
+]
+
 [tool.uv.sources]
 generals-bots = { git = "https://github.com/strakam/generals-bots.git" }
 

requirements-competition.txt

@@ -0,0 +1,11 @@
+numpy==2.4.6
+scipy==1.18.0
+pandas==3.0.5
+scikit-learn==1.9.0
+# On Linux, the default PyPI wheel bundles CUDA. For a CPU-only build, uncomment:
+jax==0.11.0
+numba==0.66.0
+networkx==3.6.1
+safetensors==0.8.0
+gymnasium==1.3.0
+# torch==2.13.0

requirements-vendor.txt

@@ -0,0 +1,3 @@
+equinox==0.13.8
+jaxtyping==0.3.11
+wadler-lindig==0.1.7

requirements.txt

@@ -0,0 +1,81 @@
+# This file was autogenerated by uv via the following command:
+#    uv pip compile pyproject.toml -o requirements.txt
+absl-py==2.5.0
+    # via optax
+antlr4-python3-runtime==4.9.3
+    # via omegaconf
+bidict==0.23.1
+    # via python-socketio
+certifi==2026.7.22
+    # via requests
+charset-normalizer==3.4.9
+    # via requests
+equinox==0.13.8
+    # via general-bots-training (pyproject.toml)
+generals-bots @ git+https://github.com/strakam/generals-bots.git@9e3b9d13cca51caa1bb07db48bb85c9e90ce0462
+    # via general-bots-training (pyproject.toml)
+h11==0.16.0
+    # via wsproto
+idna==3.18
+    # via requests
+jax==0.11.0
+    # via
+    #   general-bots-training (pyproject.toml)
+    #   equinox
+    #   generals-bots
+    #   optax
+jaxlib==0.11.0
+    # via
+    #   generals-bots
+    #   jax
+    #   optax
+jaxtyping==0.3.11
+    # via
+    #   general-bots-training (pyproject.toml)
+    #   equinox
+ml-dtypes==0.5.4
+    # via
+    #   jax
+    #   jaxlib
+numpy==2.5.1
+    # via
+    #   generals-bots
+    #   jax
+    #   jaxlib
+    #   ml-dtypes
+    #   optax
+    #   scipy
+omegaconf==2.3.1
+    # via general-bots-training (pyproject.toml)
+opt-einsum==3.4.0
+    # via jax
+optax==0.2.8
+    # via general-bots-training (pyproject.toml)
+pygame==2.6.1
+    # via generals-bots
+python-engineio==4.13.4
+    # via python-socketio
+python-socketio==5.16.4
+    # via generals-bots
+pyyaml==6.0.3
+    # via omegaconf
+requests==2.34.2
+    # via python-socketio
+scipy==1.18.0
+    # via
+    #   jax
+    #   jaxlib
+simple-websocket==1.1.0
+    # via python-engineio
+typing-extensions==4.16.0
+    # via equinox
+urllib3==2.7.0
+    # via requests
+wadler-lindig==0.1.7
+    # via
+    #   equinox
+    #   jaxtyping
+websocket-client==1.9.0
+    # via python-socketio
+wsproto==1.3.2
+    # via simple-websocket

scripts/competition.py

@@ -0,0 +1,342 @@
+"""Competition stdio agent: direct policy sampling over a trained equinox checkpoint.
+
+Self-contained competition bot. It speaks the wire protocol in
+`generals-bots/competition/protocol.py` (handshake, then one observation frame
+per turn, one action line per turn, EOF on stdin = game over) and, on each
+turn, samples an action from the policy network exactly as the model is used
+during training/evaluation — no tree search.
+
+Unlike `scripts/mcts_agent.py`, this module does **not** import the
+`general_bots_training` training package. It inlines the policy-value network
+and the observation encoding so the script can be dropped into the competition
+sandbox with just the engine (`generals`), `equinox`, `jax`, and `numpy`
+available.
+
+This is the search-free counterpart to `scripts/competition_puct.py`: it shares
+the same network and observation encoding, but selects an action by sampling
+from the masked policy (as in `PolicyValueNetwork.__call__`) instead of running
+particle PUCT.
+
+Run it directly through the bundled matchup driver, e.g.:
+
+    PYTHONPATH=src:generals-bots .venv/bin/python \
+      generals-bots/competition/matchup.py \
+      scripts/competition.py --checkpoint ppo_model.eqx \
+      generals-bots/competition/agents/expander_python/run.sh \
+      --mode competition
+
+or wrap it in a `run.sh` that passes the checkpoint path.
+"""
+
+import argparse
+import os
+import sys
+import time
+
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+import jax.random as jrandom
+import numpy as np
+from generals.core import game
+from generals.core.action import compute_valid_move_mask
+from generals.core.game import GameState
+from generals.core.observation import Observation
+from generals.modifiers import build_castles, deathtouch
+
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+
+PASS_ACTION = np.array([1, 0, 0, 0, 0], dtype=np.int32)
+
+
+# ---------------------------------------------------------------------------
+# Policy-value network (inlined from general_bots_training.network)
+# ---------------------------------------------------------------------------
+
+
+def obs_to_tensor(obs: Observation) -> jnp.ndarray:
+    """Encode an Observation into a (C, H, W) float32 tensor for the network.
+
+    Armies and army-counts are log-normalized; land counts are divided by the
+    number of cells; scalar values are broadcast to spatial planes so a plain
+    conv stack can consume them. Grid-size agnostic.
+    """
+    H, W = obs.armies.shape
+    armies = jnp.log1p(obs.armies.astype(jnp.float32)) / jnp.log(50.0)
+
+    def bcast(scalar):
+        return jnp.broadcast_to(scalar.astype(jnp.float32), (H, W))
+
+    own_land = bcast(obs.owned_land_count / (H * W))
+    own_army = bcast(jnp.log1p(obs.owned_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    opp_land = bcast(obs.opponent_land_count / (H * W))
+    opp_army = bcast(jnp.log1p(obs.opponent_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    timestep = bcast(jnp.log1p(obs.timestep.astype(jnp.float32)) / jnp.log(1201.0))
+
+    return jnp.stack(
+        [
+            armies,
+            obs.generals.astype(jnp.float32),
+            obs.castles.astype(jnp.float32),
+            obs.mountains.astype(jnp.float32),
+            obs.neutral_cells.astype(jnp.float32),
+            obs.owned_cells.astype(jnp.float32),
+            obs.opponent_cells.astype(jnp.float32),
+            obs.fog_cells.astype(jnp.float32),
+            obs.structures_in_fog.astype(jnp.float32),
+            own_land,
+            own_army,
+            opp_land,
+            opp_army,
+            timestep,
+        ],
+        axis=0,
+    )
+
+
+class PolicyValueNetwork(eqx.Module):
+    """Conv policy-value network.
+
+    Action layout: 4 full-move directions and 4 half-move (split) directions
+    per source cell, plus one global pass action. Invalid moves are masked to
+    -1e9; pass is always available.
+
+    The value head uses global average pooling over the spatial grid, so the
+    network works for any board size without reshaping linear layers.
+    """
+
+    conv1: eqx.nn.Conv2d
+    conv2: eqx.nn.Conv2d
+    conv3: eqx.nn.Conv2d
+    conv4: eqx.nn.Conv2d
+    policy_conv: eqx.nn.Conv2d
+    value_conv: eqx.nn.Conv2d
+    value_linear1: eqx.nn.Linear
+    value_linear2: eqx.nn.Linear
+
+    def __init__(self, key, in_channels: int = 14, channels=(32, 32, 32, 16)):
+        keys = jrandom.split(key, 8)
+        self.conv1 = eqx.nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1, key=keys[0])
+        self.conv2 = eqx.nn.Conv2d(channels[0], channels[1], kernel_size=3, padding=1, key=keys[1])
+        self.conv3 = eqx.nn.Conv2d(channels[1], channels[2], kernel_size=3, padding=1, key=keys[2])
+        self.conv4 = eqx.nn.Conv2d(channels[2], channels[3], kernel_size=3, padding=1, key=keys[3])
+        # 9 = 4 dirs (full) + 4 dirs (half) + 1 pass
+        self.policy_conv = eqx.nn.Conv2d(channels[3], 9, kernel_size=1, key=keys[4])
+        self.value_conv = eqx.nn.Conv2d(channels[3], 4, kernel_size=1, key=keys[5])
+        self.value_linear1 = eqx.nn.Linear(4, 64, key=keys[6])
+        self.value_linear2 = eqx.nn.Linear(64, 1, key=keys[7])
+
+    def _features(self, obs):
+        x = jax.nn.relu(self.conv1(obs))
+        x = jax.nn.relu(self.conv2(x))
+        x = jax.nn.relu(self.conv3(x))
+        x = jax.nn.relu(self.conv4(x))
+        return x
+
+    def _value_from_features(self, feat):
+        v = jax.nn.relu(self.value_conv(feat))  # (4, H, W)
+        v = v.mean(axis=(1, 2))  # (4,) global average pool
+        v = jax.nn.relu(self.value_linear1(v))
+        return self.value_linear2(v)[0]
+
+    def policy_value(self, obs, mask):
+        """Return deterministic masked policy logits and the critic value."""
+        features = self._features(obs)
+        return self._policy_logits(features, mask), self._value_from_features(features)
+
+    def _policy_logits(self, feat, mask):
+        """Return 8*H*W move logits plus one global pass logit."""
+        logits = self.policy_conv(feat)  # (9, H, W)
+        mask_t = jnp.transpose(mask, (2, 0, 1))  # (4, H, W)
+        penalty = (1.0 - mask_t) * -1e9
+        move_penalty = jnp.concatenate([penalty, penalty], axis=0)
+        move_logits = (logits[:8] + move_penalty).reshape(-1)
+        pass_logit = jnp.mean(logits[8])[None]
+        return jnp.concatenate([move_logits, pass_logit])
+
+
+# ---------------------------------------------------------------------------
+# Action selection (matches training/evaluation rollout)
+# ---------------------------------------------------------------------------
+
+
+@jax.jit
+def competition_move_mask(obs: Observation) -> jnp.ndarray:
+    """Legal move mask, identical to the one used during training/evaluation."""
+    return compute_valid_move_mask(obs.armies, obs.owned_cells, obs.mountains)
+
+
+def _decode_policy_index(index: int, height: int, width: int) -> np.ndarray:
+    cells = height * width
+    if index == 8 * cells:
+        return PASS_ACTION.copy()
+    encoded_direction, position = divmod(index, cells)
+    row, col = divmod(position, width)
+    split = int(encoded_direction >= 4)
+    direction = encoded_direction - 4 if split else encoded_direction
+    return np.array([0, row, col, direction, split], dtype=np.int32)
+
+
+def observation_from_wire(
+    timestep: int,
+    owned_land_count: int,
+    owned_army_count: int,
+    opponent_land_count: int,
+    opponent_army_count: int,
+    type_grid,
+    owner_grid,
+    army_grid,
+) -> Observation:
+    """Convert a competition wire frame into the network's Observation type."""
+    types = jnp.asarray(type_grid, dtype=jnp.int32)
+    owners = jnp.asarray(owner_grid, dtype=jnp.int32)
+    armies = jnp.asarray(army_grid, dtype=jnp.int32)
+    visible = (types != 0) & (types != 5)
+    return Observation(
+        armies=armies,
+        generals=types == 4,
+        castles=types == 3,
+        mountains=types == 2,
+        neutral_cells=visible & (owners == 0) & (types != 2),
+        owned_cells=owners == 1,
+        opponent_cells=owners == 2,
+        fog_cells=types == 0,
+        structures_in_fog=types == 5,
+        owned_land_count=jnp.int32(owned_land_count),
+        owned_army_count=jnp.int32(owned_army_count),
+        opponent_land_count=jnp.int32(opponent_land_count),
+        opponent_army_count=jnp.int32(opponent_army_count),
+        timestep=jnp.int32(timestep),
+    )
+
+
+@jax.jit
+def competition_step(state: GameState, actions: jnp.ndarray):
+    state, actions = build_castles.apply_build_actions(state, actions)
+    return deathtouch.step(state, actions, turn=800)
+
+
+class PolicyAgent:
+    """Direct policy agent that samples actions as in training/evaluation.
+
+    On each turn it computes the masked policy logits over all legal moves and
+    samples one action with `jrandom.categorical`, exactly mirroring
+    `PolicyValueNetwork.__call__` used during rollout collection.
+    """
+
+    def __init__(
+        self,
+        network: PolicyValueNetwork,
+        seed: int = 0,
+    ):
+        self.network = network
+        self.key = jrandom.PRNGKey(seed)
+        self._infer = eqx.filter_jit(
+            lambda network, tensor, mask: network.policy_value(tensor, mask)
+        )
+
+    def warmup(self, height: int, width: int) -> None:
+        """Compile shape-dependent network and competition transition kernels."""
+        grid = jnp.zeros((height, width), dtype=jnp.int32)
+        grid = grid.at[0, 0].set(1).at[height - 1, width - 1].set(2)
+        state = game.create_initial_state(grid)
+        state = state._replace(
+            armies=state.armies.at[0, 0].set(100).at[height - 1, width - 1].set(100),
+            time=jnp.int32(100),
+        )
+        obs = game.get_observation(state, 0)
+        self.policy_value(obs)
+        actions = jnp.stack([jnp.asarray(PASS_ACTION), jnp.asarray(PASS_ACTION)])
+        warmed_state, _ = competition_step(state, actions)
+        jax.block_until_ready(warmed_state)
+
+    def policy_value(self, obs: Observation):
+        mask = competition_move_mask(obs)
+        logits, value = self._infer(self.network, obs_to_tensor(obs), mask)
+        return np.asarray(logits), float(value), mask
+
+    def select(self, obs: Observation) -> tuple[np.ndarray, dict[str, float]]:
+        """Sample one action from the masked policy and report statistics."""
+        started_at = time.perf_counter()
+        logits, value, _ = self.policy_value(obs)
+        self.key, sample_key = jrandom.split(self.key)
+        index = int(jrandom.categorical(sample_key, jnp.asarray(logits)))
+        height, width = obs.armies.shape
+        action = _decode_policy_index(index, height, width)
+        elapsed_ms = (time.perf_counter() - started_at) * 1000
+        return action, {
+            "value": value,
+            "elapsed_ms": elapsed_ms,
+        }
+
+
+# ---------------------------------------------------------------------------
+# Stdio driver
+# ---------------------------------------------------------------------------
+
+
+def _read_grid(stream, height: int):
+    return [[int(value) for value in stream.readline().split()] for _ in range(height)]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("checkpoint", help="Equinox policy checkpoint")
+    # Accepted for compatibility with the original run.sh; the direct policy
+    # agent does no search, so these are ignored.
+    parser.add_argument("--time-budget-ms", type=float, default=125.0)
+    parser.add_argument("--max-simulations", type=int, default=128)
+    parser.add_argument("--rollout-depth", type=int, default=2)
+    parser.add_argument("--top-k", type=int, default=20)
+    parser.add_argument("--seed", type=int, default=0)
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    handshake = sys.stdin.readline()
+    if not handshake:
+        return
+    player_index, height, width = (int(value) for value in handshake.split())
+
+    network = PolicyValueNetwork(jrandom.PRNGKey(args.seed))
+    network = eqx.tree_deserialise_leaves(args.checkpoint, network)
+    agent = PolicyAgent(network, seed=args.seed)
+    agent.warmup(height, width)
+    print(f"[competition] warmup complete for {height}x{width}", file=sys.stderr, flush=True)
+
+    while True:
+        scalar_line = sys.stdin.readline()
+        if not scalar_line:
+            return
+        timestep, own_land, own_army, opponent_land, opponent_army = (
+            int(value) for value in scalar_line.split()
+        )
+        type_grid = _read_grid(sys.stdin, height)
+        owner_grid = _read_grid(sys.stdin, height)
+        army_grid = _read_grid(sys.stdin, height)
+        observation = observation_from_wire(
+            timestep,
+            own_land,
+            own_army,
+            opponent_land,
+            opponent_army,
+            type_grid,
+            owner_grid,
+            army_grid,
+        )
+
+        started_at = time.perf_counter()
+        action, stats = agent.select(observation)
+        elapsed_ms = (time.perf_counter() - started_at) * 1000
+        print(
+            f"[competition] turn={timestep} value={stats['value']:+.3f} "
+            f"elapsed_ms={elapsed_ms:.1f}",
+            file=sys.stderr,
+            flush=True,
+        )
+        print(" ".join(str(int(value)) for value in action), flush=True)
+
+
+if __name__ == "__main__":
+    main()

scripts/competition_puct.py

@@ -0,0 +1,552 @@
+"""Competition stdio agent: particle PUCT over a trained equinox checkpoint.
+
+Self-contained competition bot. It speaks the wire protocol in
+`generals-bots/competition/protocol.py` (handshake, then one observation frame
+per turn, one action line per turn, EOF on stdin = game over) and runs
+deadline-bounded root PUCT using only the perspective-relative wire
+observation.
+
+Unlike `scripts/mcts_agent.py`, this module does **not** import the
+`general_bots_training` training package. It inlines the policy-value network,
+the observation encoding, and the particle search so the script can be dropped
+into the competition sandbox with just the engine (`generals`), `equinox`,
+`jax`, and `numpy` available.
+
+Run it directly through the bundled matchup driver, e.g.:
+
+    PYTHONPATH=src:generals-bots .venv/bin/python \
+      generals-bots/competition/matchup.py \
+      scripts/competition.py --checkpoint ppo_model.eqx \
+      generals-bots/competition/agents/expander_python/run.sh \
+      --mode competition
+
+or wrap it in a `run.sh` that passes the checkpoint path.
+"""
+
+import argparse
+import os
+import sys
+import time
+from dataclasses import dataclass
+
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+import jax.random as jrandom
+import numpy as np
+from generals.core import game
+from generals.core.action import compute_valid_move_mask
+from generals.core.game import GameState
+from generals.core.observation import Observation
+from generals.modifiers import build_castles, deathtouch
+
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+
+PASS_ACTION = np.array([1, 0, 0, 0, 0], dtype=np.int32)
+
+# Search/agent defaults, previously exposed as CLI flags. Kept as constants so
+# the script can be launched by a bare `run.sh` with no extra arguments.
+TIME_BUDGET_MS = 125.0
+MAX_SIMULATIONS = 128
+ROLLOUT_DEPTH = 2
+TOP_K = 20
+SEED = 0
+
+
+# ---------------------------------------------------------------------------
+# Policy-value network (inlined from general_bots_training.network)
+# ---------------------------------------------------------------------------
+
+
+def obs_to_tensor(obs: Observation) -> jnp.ndarray:
+    """Encode an Observation into a (C, H, W) float32 tensor for the network.
+
+    Armies and army-counts are log-normalized; land counts are divided by the
+    number of cells; scalar values are broadcast to spatial planes so a plain
+    conv stack can consume them. Grid-size agnostic.
+    """
+    H, W = obs.armies.shape
+    armies = jnp.log1p(obs.armies.astype(jnp.float32)) / jnp.log(50.0)
+
+    def bcast(scalar):
+        return jnp.broadcast_to(scalar.astype(jnp.float32), (H, W))
+
+    own_land = bcast(obs.owned_land_count / (H * W))
+    own_army = bcast(jnp.log1p(obs.owned_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    opp_land = bcast(obs.opponent_land_count / (H * W))
+    opp_army = bcast(jnp.log1p(obs.opponent_army_count.astype(jnp.float32)) / jnp.log(50.0))
+    timestep = bcast(jnp.log1p(obs.timestep.astype(jnp.float32)) / jnp.log(1201.0))
+
+    return jnp.stack(
+        [
+            armies,
+            obs.generals.astype(jnp.float32),
+            obs.castles.astype(jnp.float32),
+            obs.mountains.astype(jnp.float32),
+            obs.neutral_cells.astype(jnp.float32),
+            obs.owned_cells.astype(jnp.float32),
+            obs.opponent_cells.astype(jnp.float32),
+            obs.fog_cells.astype(jnp.float32),
+            obs.structures_in_fog.astype(jnp.float32),
+            own_land,
+            own_army,
+            opp_land,
+            opp_army,
+            timestep,
+        ],
+        axis=0,
+    )
+
+
+class PolicyValueNetwork(eqx.Module):
+    """Conv policy-value network.
+
+    Action layout: 4 full-move directions and 4 half-move (split) directions
+    per source cell, plus one global pass action. Invalid moves are masked to
+    -1e9; pass is always available.
+
+    The value head uses global average pooling over the spatial grid, so the
+    network works for any board size without reshaping linear layers.
+    """
+
+    conv1: eqx.nn.Conv2d
+    conv2: eqx.nn.Conv2d
+    conv3: eqx.nn.Conv2d
+    conv4: eqx.nn.Conv2d
+    policy_conv: eqx.nn.Conv2d
+    value_conv: eqx.nn.Conv2d
+    value_linear1: eqx.nn.Linear
+    value_linear2: eqx.nn.Linear
+
+    def __init__(self, key, in_channels: int = 14, channels=(32, 32, 32, 16)):
+        keys = jrandom.split(key, 8)
+        self.conv1 = eqx.nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1, key=keys[0])
+        self.conv2 = eqx.nn.Conv2d(channels[0], channels[1], kernel_size=3, padding=1, key=keys[1])
+        self.conv3 = eqx.nn.Conv2d(channels[1], channels[2], kernel_size=3, padding=1, key=keys[2])
+        self.conv4 = eqx.nn.Conv2d(channels[2], channels[3], kernel_size=3, padding=1, key=keys[3])
+        # 9 = 4 dirs (full) + 4 dirs (half) + 1 pass
+        self.policy_conv = eqx.nn.Conv2d(channels[3], 9, kernel_size=1, key=keys[4])
+        self.value_conv = eqx.nn.Conv2d(channels[3], 4, kernel_size=1, key=keys[5])
+        self.value_linear1 = eqx.nn.Linear(4, 64, key=keys[6])
+        self.value_linear2 = eqx.nn.Linear(64, 1, key=keys[7])
+
+    def _features(self, obs):
+        x = jax.nn.relu(self.conv1(obs))
+        x = jax.nn.relu(self.conv2(x))
+        x = jax.nn.relu(self.conv3(x))
+        x = jax.nn.relu(self.conv4(x))
+        return x
+
+    def _value_from_features(self, feat):
+        v = jax.nn.relu(self.value_conv(feat))  # (4, H, W)
+        v = v.mean(axis=(1, 2))  # (4,) global average pool
+        v = jax.nn.relu(self.value_linear1(v))
+        return self.value_linear2(v)[0]
+
+    def policy_value(self, obs, mask):
+        """Return deterministic masked policy logits and the critic value."""
+        features = self._features(obs)
+        return self._policy_logits(features, mask), self._value_from_features(features)
+
+    def _policy_logits(self, feat, mask):
+        """Return 8*H*W move logits plus one global pass logit."""
+        logits = self.policy_conv(feat)  # (9, H, W)
+        mask_t = jnp.transpose(mask, (2, 0, 1))  # (4, H, W)
+        penalty = (1.0 - mask_t) * -1e9
+        move_penalty = jnp.concatenate([penalty, penalty], axis=0)
+        move_logits = (logits[:8] + move_penalty).reshape(-1)
+        pass_logit = jnp.mean(logits[8])[None]
+        return jnp.concatenate([move_logits, pass_logit])
+
+
+# ---------------------------------------------------------------------------
+# Particle PUCT search (inlined from general_bots_training.mcts)
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class MCTSConfig:
+    time_budget_ms: float = 100.0
+    max_simulations: int = 128
+    rollout_depth: int = 2
+    top_k: int = 20
+    opponent_top_k: int = 12
+    max_build_actions: int = 3
+    c_puct: float = 1.5
+    value_scale: float = 5.0
+
+
+@jax.jit
+def competition_move_mask(obs: Observation) -> jnp.ndarray:
+    """Legal move mask that conservatively blocks unresolved fog structures."""
+    blocked = obs.mountains | obs.structures_in_fog
+    return compute_valid_move_mask(obs.armies, obs.owned_cells, blocked)
+
+
+@jax.jit
+def build_cost_grid(obs: Observation) -> jnp.ndarray:
+    """Exact own-castle build costs derivable from a fog observation."""
+    structures = ((obs.castles | obs.generals) & obs.owned_cells).astype(jnp.int32)
+    height, width = structures.shape
+    radius = 6
+    padded = jnp.pad(structures, radius)
+    costs = jnp.full((height, width), 35, dtype=jnp.int32)
+    for row_offset in range(-radius, radius + 1):
+        for col_offset in range(-radius, radius + 1):
+            surcharge = 14 - 2 * (abs(row_offset) + abs(col_offset))
+            if surcharge > 0:
+                shifted = padded[
+                    radius + row_offset : radius + row_offset + height,
+                    radius + col_offset : radius + col_offset + width,
+                ]
+                costs = costs + surcharge * shifted
+    return costs
+
+
+@jax.jit
+def valid_build_mask(obs: Observation) -> jnp.ndarray:
+    costs = build_cost_grid(obs)
+    plain_owned = obs.owned_cells & ~obs.generals & ~obs.castles
+    return plain_owned & (obs.armies >= costs)
+
+
+def _decode_policy_index(index: int, height: int, width: int) -> np.ndarray:
+    cells = height * width
+    if index == 8 * cells:
+        return PASS_ACTION.copy()
+    encoded_direction, position = divmod(index, cells)
+    row, col = divmod(position, width)
+    split = int(encoded_direction >= 4)
+    direction = encoded_direction - 4 if split else encoded_direction
+    return np.array([0, row, col, direction, split], dtype=np.int32)
+
+
+def observation_from_wire(
+    timestep: int,
+    owned_land_count: int,
+    owned_army_count: int,
+    opponent_land_count: int,
+    opponent_army_count: int,
+    type_grid,
+    owner_grid,
+    army_grid,
+) -> Observation:
+    """Convert a competition wire frame into the network's Observation type."""
+    types = jnp.asarray(type_grid, dtype=jnp.int32)
+    owners = jnp.asarray(owner_grid, dtype=jnp.int32)
+    armies = jnp.asarray(army_grid, dtype=jnp.int32)
+    visible = (types != 0) & (types != 5)
+    return Observation(
+        armies=armies,
+        generals=types == 4,
+        castles=types == 3,
+        mountains=types == 2,
+        neutral_cells=visible & (owners == 0) & (types != 2),
+        owned_cells=owners == 1,
+        opponent_cells=owners == 2,
+        fog_cells=types == 0,
+        structures_in_fog=types == 5,
+        owned_land_count=jnp.int32(owned_land_count),
+        owned_army_count=jnp.int32(owned_army_count),
+        opponent_land_count=jnp.int32(opponent_land_count),
+        opponent_army_count=jnp.int32(opponent_army_count),
+        timestep=jnp.int32(timestep),
+    )
+
+
+def sample_determinization(
+    obs: Observation, player_index: int, rng: np.random.Generator
+) -> GameState:
+    """Sample a conservative hidden state consistent with visible cells/totals."""
+    armies = np.asarray(obs.armies, dtype=np.int32).copy()
+    owned = np.asarray(obs.owned_cells, dtype=bool)
+    visible_opponent = np.asarray(obs.opponent_cells, dtype=bool)
+    mountains = np.asarray(obs.mountains | obs.structures_in_fog, dtype=bool)
+    castles = np.asarray(obs.castles, dtype=bool)
+    generals = np.asarray(obs.generals, dtype=bool).copy()
+    fog = np.asarray(obs.fog_cells, dtype=bool) & ~mountains
+
+    opponent = visible_opponent.copy()
+    hidden_candidates = np.argwhere(fog & ~owned)
+    missing_land = max(0, int(obs.opponent_land_count) - int(opponent.sum()))
+    if len(hidden_candidates):
+        chosen = hidden_candidates[
+            rng.choice(
+                len(hidden_candidates),
+                size=min(missing_land, len(hidden_candidates)),
+                replace=False,
+            )
+        ]
+        opponent[chosen[:, 0], chosen[:, 1]] = True
+
+    visible_enemy_general = generals & opponent
+    if not visible_enemy_general.any():
+        candidates = np.argwhere((opponent | fog) & ~owned & ~mountains)
+        if len(candidates):
+            row, col = candidates[rng.integers(len(candidates))]
+            generals[row, col] = True
+            opponent[row, col] = True
+
+    hidden_opponent = opponent & ~visible_opponent
+    remaining_army = max(0, int(obs.opponent_army_count) - int(armies[visible_opponent].sum()))
+    hidden_cells = np.argwhere(hidden_opponent)
+    if len(hidden_cells):
+        base = min(remaining_army, len(hidden_cells))
+        armies[hidden_opponent] = 0
+        armies[hidden_cells[:base, 0], hidden_cells[:base, 1]] = 1
+        remaining_army -= base
+        if remaining_army:
+            allocations = rng.multinomial(
+                remaining_army, np.full(len(hidden_cells), 1 / len(hidden_cells))
+            )
+            armies[hidden_cells[:, 0], hidden_cells[:, 1]] += allocations.astype(np.int32)
+
+    ownership_relative = np.stack([owned, opponent])
+    if player_index == 1:
+        ownership = ownership_relative[::-1]
+    else:
+        ownership = ownership_relative
+    passable = ~mountains
+    neutral = passable & ~ownership[0] & ~ownership[1]
+
+    general_positions = []
+    for absolute_player in range(2):
+        positions = np.argwhere(generals & ownership[absolute_player])
+        if len(positions):
+            general_positions.append(positions[0])
+        else:
+            fallback = np.argwhere(ownership[absolute_player] & passable)
+            general_positions.append(fallback[0] if len(fallback) else np.array([0, 0]))
+
+    return GameState(
+        armies=jnp.asarray(armies),
+        ownership=jnp.asarray(ownership),
+        ownership_neutral=jnp.asarray(neutral),
+        generals=jnp.asarray(generals),
+        castles=jnp.asarray(castles),
+        mountains=jnp.asarray(mountains),
+        passable=jnp.asarray(passable),
+        general_positions=jnp.asarray(general_positions, dtype=jnp.int32),
+        time=jnp.asarray(obs.timestep, dtype=jnp.int32),
+        winner=jnp.int32(-1),
+        pool_idx=jnp.int32(0),
+    )
+
+
+@jax.jit
+def competition_step(state: GameState, actions: jnp.ndarray):
+    state, actions = build_castles.apply_build_actions(state, actions)
+    return deathtouch.step(state, actions, turn=800)
+
+
+class ParticlePUCT:
+    """Deadline-bounded root PUCT with policy-guided simultaneous rollouts."""
+
+    def __init__(
+        self,
+        network: PolicyValueNetwork,
+        player_index: int,
+        config: MCTSConfig = MCTSConfig(),
+        seed: int = 0,
+    ):
+        self.network = network
+        self.player_index = player_index
+        self.config = config
+        self.rng = np.random.default_rng(seed)
+        self._infer = eqx.filter_jit(
+            lambda network, tensor, mask: network.policy_value(tensor, mask)
+        )
+
+    def warmup(self, height: int, width: int) -> None:
+        """Compile shape-dependent network and competition transition kernels."""
+        grid = jnp.zeros((height, width), dtype=jnp.int32)
+        grid = grid.at[0, 0].set(1).at[height - 1, width - 1].set(2)
+        state = game.create_initial_state(grid)
+        state = state._replace(
+            armies=state.armies.at[0, 0].set(100).at[height - 1, width - 1].set(100),
+            time=jnp.int32(100),
+        )
+        obs = game.get_observation(state, self.player_index)
+        self.policy_value(obs)
+        actions = jnp.stack([jnp.asarray(PASS_ACTION), jnp.asarray(PASS_ACTION)])
+        warmed_state, _ = competition_step(state, actions)
+        jax.block_until_ready(warmed_state)
+        candidates, _ = self.candidates(obs)
+        self._simulate(obs, candidates[0])
+        self._simulate(obs, candidates[min(1, len(candidates) - 1)])
+
+    def policy_value(self, obs: Observation):
+        mask = competition_move_mask(obs)
+        logits, value = self._infer(self.network, obs_to_tensor(obs), mask)
+        return np.asarray(logits), float(value), mask
+
+    def candidates(self, obs: Observation, top_k: int | None = None):
+        logits, _, _ = self.policy_value(obs)
+        height, width = obs.armies.shape
+        valid_indices = np.flatnonzero(logits > -1e8)
+        count = min(top_k or self.config.top_k, len(valid_indices))
+        ranked = np.argsort(logits[valid_indices])[::-1][:count]
+        indices = valid_indices[ranked]
+        selected_logits = logits[indices]
+        selected_logits = selected_logits - selected_logits.max()
+        priors = np.exp(selected_logits)
+        actions = [_decode_policy_index(int(index), height, width) for index in indices]
+
+        build_mask = np.asarray(valid_build_mask(obs))
+        build_positions = np.argwhere(build_mask)
+        if len(build_positions):
+            costs = np.asarray(build_cost_grid(obs))
+            armies = np.asarray(obs.armies)
+            scores = np.array([armies[r, c] - costs[r, c] for r, c in build_positions])
+            order = np.argsort(scores)[::-1][: self.config.max_build_actions]
+            build_prior = max(float(priors.sum()) * 0.05, 1e-3)
+            for position_index in order:
+                row, col = build_positions[position_index]
+                actions.append(np.array([2, row, col, 0, 0], dtype=np.int32))
+                priors = np.append(priors, build_prior)
+
+        priors = priors / priors.sum()
+        return actions, priors
+
+    def _sample_policy_action(self, obs: Observation, top_k: int) -> np.ndarray:
+        actions, priors = self.candidates(obs, top_k)
+        return actions[int(self.rng.choice(len(actions), p=priors))]
+
+    def _leaf_value(self, state: GameState, info) -> float:
+        if bool(info.is_done):
+            winner = int(info.winner)
+            return 0.0 if winner < 0 else (1.0 if winner == self.player_index else -1.0)
+        obs = game.get_observation(state, self.player_index)
+        _, value, _ = self.policy_value(obs)
+        return float(np.tanh(value / self.config.value_scale))
+
+    def _simulate(self, root_obs: Observation, root_action: np.ndarray) -> float:
+        state = sample_determinization(root_obs, self.player_index, self.rng)
+        info = game.get_info(state)
+        our_action = root_action
+        for _ in range(self.config.rollout_depth):
+            opponent_index = 1 - self.player_index
+            opponent_obs = game.get_observation(state, opponent_index)
+            opponent_action = self._sample_policy_action(opponent_obs, self.config.opponent_top_k)
+            joint_actions = [None, None]
+            joint_actions[self.player_index] = our_action
+            joint_actions[opponent_index] = opponent_action
+            state, info = competition_step(state, jnp.asarray(joint_actions, dtype=jnp.int32))
+            jax.block_until_ready(state)
+            if bool(info.is_done):
+                break
+            our_obs = game.get_observation(state, self.player_index)
+            our_action = self._sample_policy_action(our_obs, self.config.top_k)
+        return self._leaf_value(state, info)
+
+    def search(self, obs: Observation) -> tuple[np.ndarray, dict[str, float]]:
+        started_at = time.perf_counter()
+        deadline = started_at + self.config.time_budget_ms / 1000.0
+        actions, priors = self.candidates(obs)
+        visits = np.zeros(len(actions), dtype=np.int32)
+        value_sums = np.zeros(len(actions), dtype=np.float64)
+        simulations = 0
+        estimated_simulation_seconds = 0.0
+
+        while (
+            simulations < self.config.max_simulations
+            and time.perf_counter() + estimated_simulation_seconds < deadline
+        ):
+            total_visits = max(1, int(visits.sum()))
+            q_values = np.divide(
+                value_sums,
+                visits,
+                out=np.zeros_like(value_sums),
+                where=visits > 0,
+            )
+            scores = q_values + self.config.c_puct * priors * np.sqrt(total_visits) / (1 + visits)
+            action_index = int(np.argmax(scores))
+            simulation_started_at = time.perf_counter()
+            value = self._simulate(obs, actions[action_index])
+            simulation_elapsed = time.perf_counter() - simulation_started_at
+            estimated_simulation_seconds = max(estimated_simulation_seconds, simulation_elapsed)
+            visits[action_index] += 1
+            value_sums[action_index] += value
+            simulations += 1
+
+        selected = int(np.argmax(visits)) if simulations else int(np.argmax(priors))
+        return actions[selected], {
+            "simulations": float(simulations),
+            "selected_visits": float(visits[selected]),
+            "elapsed_budget_ms": self.config.time_budget_ms,
+        }
+
+
+# ---------------------------------------------------------------------------
+# Stdio driver
+# ---------------------------------------------------------------------------
+
+
+def _read_grid(stream, height: int):
+    return [[int(value) for value in stream.readline().split()] for _ in range(height)]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("checkpoint", help="Equinox policy checkpoint")
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    handshake = sys.stdin.readline()
+    if not handshake:
+        return
+    player_index, height, width = (int(value) for value in handshake.split())
+
+    network = PolicyValueNetwork(jrandom.PRNGKey(SEED))
+    network = eqx.tree_deserialise_leaves(args.checkpoint, network)
+    search = ParticlePUCT(
+        network,
+        player_index,
+        MCTSConfig(
+            time_budget_ms=TIME_BUDGET_MS,
+            max_simulations=MAX_SIMULATIONS,
+            rollout_depth=ROLLOUT_DEPTH,
+            top_k=TOP_K,
+        ),
+        seed=SEED,
+    )
+    search.warmup(height, width)
+    print(f"[competition] warmup complete for {height}x{width}", file=sys.stderr, flush=True)
+
+    while True:
+        scalar_line = sys.stdin.readline()
+        if not scalar_line:
+            return
+        timestep, own_land, own_army, opponent_land, opponent_army = (
+            int(value) for value in scalar_line.split()
+        )
+        type_grid = _read_grid(sys.stdin, height)
+        owner_grid = _read_grid(sys.stdin, height)
+        army_grid = _read_grid(sys.stdin, height)
+        observation = observation_from_wire(
+            timestep,
+            own_land,
+            own_army,
+            opponent_land,
+            opponent_army,
+            type_grid,
+            owner_grid,
+            army_grid,
+        )
+
+        started_at = time.perf_counter()
+        action, stats = search.search(observation)
+        elapsed_ms = (time.perf_counter() - started_at) * 1000
+        print(
+            f"[competition] turn={timestep} simulations={int(stats['simulations'])} "
+            f"elapsed_ms={elapsed_ms:.1f}",
+            file=sys.stderr,
+            flush=True,
+        )
+        print(" ".join(str(int(value)) for value in action), flush=True)
+
+
+if __name__ == "__main__":
+    main()

scripts/evaluate.py

@@ -0,0 +1,123 @@
+"""Evaluate Generals.io agents by playing them against each other.
+
+Run with `.venv/bin/python scripts/evaluate.py` or `uv run python scripts/evaluate.py`.
+Override defaults with OmegaConf arguments such as `num_games=200 agent0=hunter`.
+
+Each agent is either one of the bundled JAX agents (`random`, `expander`,
+`hunter`) or a trained policy loaded from an equinox checkpoint via
+`agent0=model:ppo_model.eqx`. Games are run in a single vmapped batch of
+`num_games` parallel envs.
+"""
+
+from dataclasses import dataclass
+
+import jax
+import jax.numpy as jnp
+import jax.random as jrandom
+from generals.core import game
+from generals.core.env import GeneralsEnv
+from omegaconf import DictConfig, OmegaConf
+
+from general_bots_training.opponents import make_opponent
+
+
+@dataclass
+class EvalConfig:
+    grid_dims: tuple[int, int] = (21, 21)
+    truncation: int = 1200
+    num_games: int = 100
+    agent0: str = "random"
+    agent1: str = "expander"
+    seed: int = 0
+
+
+def load_config(args: list[str] | None = None) -> DictConfig:
+    """Merge CLI overrides into the typed default evaluation configuration."""
+    defaults = OmegaConf.structured(EvalConfig)
+    return OmegaConf.merge(defaults, OmegaConf.from_cli(args))  # type: ignore
+
+
+def main(config: DictConfig):
+    key = jrandom.PRNGKey(config.seed)
+    key, env_key = jrandom.split(key, 2)
+
+    agent0 = make_opponent(config.agent0)
+    agent1 = make_opponent(config.agent1)
+    label0, label1 = config.agent0, config.agent1
+
+    env = GeneralsEnv(grid_dims=tuple(config.grid_dims), truncation=config.truncation)
+    pool, _ = env.reset(env_key)
+    get_obs = game.get_full_observation if env.perfect_info else game.get_observation
+
+    n = config.num_games
+    key, init_key = jrandom.split(key)
+    init_keys = jrandom.split(init_key, n)
+    states = jax.vmap(env.init_state)(init_keys)
+    states = states._replace(pool_idx=jnp.arange(n, dtype=states.pool_idx.dtype) % env.pool_size)
+
+    step_env = jax.vmap(env.step, in_axes=(0, 0, None))
+
+    def run_episode(states, key):
+        """Play one full game per env; return final winner per env.
+
+        The env auto-resets on done and overwrites `state.winner` with -1, so we
+        capture the winner from `timestep.info` on the first done step per env and
+        latch it for the remainder of the scan.
+        """
+
+        def body(carry, _):
+            states, latched_winner, key = carry
+            key, p0_key, p1_key = jrandom.split(key, 3)
+            obs_p0 = jax.vmap(lambda s: get_obs(s, 0))(states)
+            obs_p1 = jax.vmap(lambda s: get_obs(s, 1))(states)
+            keys_p0 = jrandom.split(p0_key, n)
+            keys_p1 = jrandom.split(p1_key, n)
+            actions_p0 = jax.vmap(agent0.act)(obs_p0, keys_p0)
+            actions_p1 = jax.vmap(agent1.act)(obs_p1, keys_p1)
+            actions = jnp.stack([actions_p0, actions_p1], axis=1)
+            timesteps, states = step_env(states, actions, pool)
+            done = timesteps.terminated | timesteps.truncated
+            # On the first done, latch the winner (truncated games stay at -1 = draw).
+            new_winner = jnp.where(
+                done & (latched_winner < 0), timesteps.info.winner, latched_winner
+            )
+            return (states, new_winner, key), done
+
+        # Run a fixed number of steps equal to the truncation length, which
+        # guarantees every env has terminated or truncated at least once.
+        (states, latched_winner, key), _ = jax.lax.scan(
+            body, (states, jnp.full(n, -1, dtype=jnp.int32), key), None, length=config.truncation
+        )
+        return latched_winner
+
+    print("Generals.io evaluation")
+    print(OmegaConf.to_yaml(config, resolve=True).rstrip())
+    print(f"device: {jax.devices()[0]}")
+    print(f"agent0: {label0}")
+    print(f"agent1: {label1}")
+    print(f"games:  {n}")
+    print()
+    print("warming up (jit compile)...")
+    winners = run_episode(states, key)
+    jax.block_until_ready(winners)
+    print("warming up done\n")
+
+    print("playing...")
+    winners = run_episode(states, key)
+    jax.block_until_ready(winners)
+
+    wins0 = int(jnp.sum(winners == 0))
+    wins1 = int(jnp.sum(winners == 1))
+    draws = int(jnp.sum(winners < 0))
+
+    print("\nsummary")
+    print("-" * 40)
+    print(f"agent0 ({label0}): {wins0:4d} wins ({wins0 / n * 100:.1f}%)")
+    print(f"agent1 ({label1}): {wins1:4d} wins ({wins1 / n * 100:.1f}%)")
+    print(f"draws:              {draws:4d}      ({draws / n * 100:.1f}%)")
+    print("-" * 40)
+    print(f"total games: {n}")
+
+
+if __name__ == "__main__":
+    main(load_config())

scripts/run.sh

@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "$(dirname "$0")"
+exec python -u competition.py ../ppo_model.eqx

src/general_bots_training/__init__.py

@@ -1,6 +1,6 @@
 from .mcts import MCTSConfig, ParticlePUCT
 from .network import PolicyValueNetwork, obs_to_tensor
-from .opponents import Opponent, SelfPlayOpponent, StaticOpponent, make_opponent
+from .opponents import ModelOpponent, Opponent, SelfPlayOpponent, StaticOpponent, make_opponent
 from .ppo import compute_advantages_and_returns, compute_gae, make_train_epoch, ppo_loss
 from .rollout import make_collect_rollout, make_rollout_step
 
@@ -17,6 +17,7 @@ __all__ = [
     "make_rollout_step",
     "make_opponent",
     "StaticOpponent",
+    "ModelOpponent",
     "SelfPlayOpponent",
     "Opponent",
 ]

src/general_bots_training/opponents.py

@@ -4,10 +4,15 @@ from typing import Protocol
 
 from dataclasses import dataclass
 
+import equinox as eqx
 import jax.numpy as jnp
+import jax.random as jrandom
 from generals.agents import Agent, ExpanderAgent, HunterAgent, RandomAgent
+from generals.core.action import compute_valid_move_mask
 from generals.core.observation import Observation
 
+from .network import PolicyValueNetwork, obs_to_tensor
+
 
 class StaticOpponent(Protocol):
     """Opponent policy that carries no per-environment mutable state."""
@@ -20,6 +25,21 @@ class SelfPlayOpponent:
     """Marker selecting the current training network as player 1."""
 
 
+class ModelOpponent:
+    """Static opponent that plays a fixed policy loaded from an equinox checkpoint."""
+
+    def __init__(self, network: PolicyValueNetwork):
+        self._network = network
+
+    def act(self, observation: Observation, key: jnp.ndarray) -> jnp.ndarray:
+        obs_arr = obs_to_tensor(observation)
+        mask = compute_valid_move_mask(
+            observation.armies, observation.owned_cells, observation.mountains
+        )
+        action, _, _, _ = self._network(obs_arr, mask, key, None)
+        return action
+
+
 Opponent = StaticOpponent | SelfPlayOpponent
 
 
@@ -30,15 +50,36 @@ OPPONENT_TYPES: dict[str, type[Agent]] = {
 }
 
 
-def make_opponent(name: str) -> Opponent:
+def make_opponent(spec: str) -> Opponent:
-    """Create an opponent strategy by configuration name."""
+    """Create an opponent strategy from a single configuration string.
+
+    The spec is either a bare opponent name ("random", "expander", "hunter",
+    "self_play") or a ``type:checkpoint`` pair for a model opponent, e.g.
+    ``"model:ppo_model.eqx"``.
+    """
+    name, sep, checkpoint = spec.partition(":")
     normalized_name = name.lower().replace("-", "_")
     if normalized_name == "self_play":
         return SelfPlayOpponent()
 
+    if normalized_name == "model":
+        if not sep:
+            raise ValueError(
+                "opponent 'model' requires a checkpoint path, e.g. 'model:ppo_model.eqx'"
+            )
+        # The key only seeds the network structure; deserialization overwrites
+        # every leaf, so a fixed seed is sufficient.
+        network = PolicyValueNetwork(jrandom.PRNGKey(0), in_channels=14)
+        network = eqx.tree_deserialise_leaves(checkpoint, network)
+        return ModelOpponent(network)
+
+    if sep:
+        choices = ", ".join([*sorted(OPPONENT_TYPES), "model", "self_play"])
+        raise ValueError(f"opponent {spec!r} does not take a checkpoint; choose one of: {choices}")
+
     try:
         opponent_type = OPPONENT_TYPES[normalized_name]
     except KeyError as error:
-        choices = ", ".join([*sorted(OPPONENT_TYPES), "self_play"])
+        choices = ", ".join([*sorted(OPPONENT_TYPES), "model", "self_play"])
-        raise ValueError(f"unknown opponent {name!r}; choose one of: {choices}") from error
+        raise ValueError(f"unknown opponent {spec!r}; choose one of: {choices}") from error
     return opponent_type()

tests/test_opponents.py

@@ -1,7 +1,7 @@
 import pytest
 from generals.agents import ExpanderAgent, HunterAgent, RandomAgent
 
-from general_bots_training.opponents import SelfPlayOpponent, make_opponent
+from general_bots_training.opponents import ModelOpponent, SelfPlayOpponent, make_opponent
 
 
 @pytest.mark.parametrize(
@@ -28,3 +28,28 @@ def test_make_opponent_supports_self_play_aliases():
 def test_make_opponent_rejects_unknown_name():
     with pytest.raises(ValueError, match="unknown opponent"):
         make_opponent("turtle")
+
+
+def test_make_opponent_model_requires_checkpoint():
+    with pytest.raises(ValueError, match="requires a checkpoint path"):
+        make_opponent("model")
+
+
+def test_make_opponent_model_loads_checkpoint(tmp_path):
+    import equinox as eqx
+    import jax.random as jrandom
+
+    from general_bots_training.network import PolicyValueNetwork
+
+    network = PolicyValueNetwork(jrandom.PRNGKey(0), in_channels=14)
+    checkpoint_path = tmp_path / "model.eqx"
+    eqx.tree_serialise_leaves(checkpoint_path, network)
+
+    opponent = make_opponent(f"model:{checkpoint_path}")
+
+    assert isinstance(opponent, ModelOpponent)
+
+
+def test_make_opponent_rejects_checkpoint_on_non_model():
+    with pytest.raises(ValueError, match="does not take a checkpoint"):
+        make_opponent("random:some.eqx")

uv.lock

@@ -132,12 +132,20 @@ source = { editable = "." }
 dependencies = [
     { name = "equinox" },
     { name = "generals-bots" },
-    { name = "jax", extra = ["cuda"] },
+    { name = "jax" },
     { name = "jaxtyping" },
     { name = "omegaconf" },
     { name = "optax" },
 ]
 
+[package.optional-dependencies]
+cuda = [
+    { name = "jax", extra = ["cuda"] },
+]
+cuda13 = [
+    { name = "jax", extra = ["cuda13"] },
+]
+
 [package.dev-dependencies]
 dev = [
     { name = "ipython" },
@@ -148,11 +156,14 @@ dev = [
 requires-dist = [
     { name = "equinox", specifier = ">=0.13.8" },
     { name = "generals-bots", git = "https://github.com/strakam/generals-bots.git" },
-    { name = "jax", extras = ["cuda"], specifier = ">=0.11.0" },
+    { name = "jax", specifier = "==0.11.0" },
+    { name = "jax", extras = ["cuda"], marker = "extra == 'cuda'", specifier = "==0.11.0" },
+    { name = "jax", extras = ["cuda13"], marker = "extra == 'cuda13'", specifier = "==0.11.0" },
     { name = "jaxtyping", specifier = ">=0.3.11" },
     { name = "omegaconf", specifier = ">=2.3.1" },
     { name = "optax", specifier = ">=0.2.8" },
 ]
+provides-extras = ["cuda", "cuda13"]
 
 [package.metadata.requires-dev]
 dev = [
@@ -253,6 +264,10 @@ cuda = [
     { name = "jax-cuda12-plugin", extra = ["with-cuda"] },
     { name = "jaxlib" },
 ]
+cuda13 = [
+    { name = "jax-cuda13-plugin", extra = ["with-cuda"] },
+    { name = "jaxlib" },
+]
 
 [[package]]
 name = "jax-cuda12-pjrt"
@@ -295,6 +310,48 @@ with-cuda = [
     { name = "nvidia-nvshmem-cu12", marker = "sys_platform == 'linux'" },
 ]
 
+[[package]]
+name = "jax-cuda13-pjrt"
+version = "0.11.0"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/dd/cd/8fa8034e2f7bc50a78c2319f576aedf06c633b94b3e5d187fb9becfc1379/jax_cuda13_pjrt-0.11.0-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:e72c7b6b266b3fbe85792b2b4569b6c87447e77c454770d5afcc081a0b87d140", size = 120632519, upload-time = "2026-07-16T18:58:54.921Z" },
+    { url = "https://files.pythonhosted.org/packages/6e/0e/140944fc5cc1e0d5891b78209fb70a2f1ce951a656292f598904ac7efe5d/jax_cuda13_pjrt-0.11.0-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:814092a06a4da93a0b01df9bcdf5f8e0900d6c0a4c3ca7e7ec4ec967de7e96d9", size = 125815692, upload-time = "2026-07-16T18:58:59.251Z" },
+]
+
+[[package]]
+name = "jax-cuda13-plugin"
+version = "0.11.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "jax-cuda13-pjrt" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/19/59/e744e3362e8139e497b1a0c1bb27666ff7b1a4a4e20007b5105afc27deaf/jax_cuda13_plugin-0.11.0-cp313-cp313-manylinux_2_27_aarch64.whl", hash = "sha256:7d11c242dc94cd24c44ce0b67f49d5771040a5e83b594b31b9ad42e88fb8821b", size = 7547142, upload-time = "2026-07-16T18:59:06.604Z" },
+    { url = "https://files.pythonhosted.org/packages/05/c2/f5eaf303b8497d3c1b8445935a8e11181591ce4a99054699d57e2cfe8f42/jax_cuda13_plugin-0.11.0-cp313-cp313-manylinux_2_27_x86_64.whl", hash = "sha256:5fda3b7b12e50b53a0d8f9cd8c6b6efe160de9b1054dd3c9652f1504d34fc9d8", size = 7571108, upload-time = "2026-07-16T18:59:08.06Z" },
+    { url = "https://files.pythonhosted.org/packages/06/14/74da5a18490a7d24bdc21dedf90b20eab4e4658879f333bea9772f36291b/jax_cuda13_plugin-0.11.0-cp314-cp314-manylinux_2_27_aarch64.whl", hash = "sha256:d9147b3db90f140bd94afb29589a7eaeb7fea1b11159420394c60d466ea9a7c1", size = 7547841, upload-time = "2026-07-16T18:59:10.03Z" },
+    { url = "https://files.pythonhosted.org/packages/42/35/9131f3182b4ef67dce4845c697c8cab79bc0df8ffb15e7806bd099cae4e3/jax_cuda13_plugin-0.11.0-cp314-cp314-manylinux_2_27_x86_64.whl", hash = "sha256:09be21c598788f0b714f27cc94e3ff8a169b7854b6db883287d6cbc6d4d8d911", size = 7571369, upload-time = "2026-07-16T18:59:11.811Z" },
+    { url = "https://files.pythonhosted.org/packages/29/67/7f057e2533f14fcdfa236deb2272e202549fd0a1b461c65d8f9c93fc2c6c/jax_cuda13_plugin-0.11.0-cp314-cp314t-manylinux_2_27_aarch64.whl", hash = "sha256:87d847502f03e7ce37e6ef570245fca1323af9a7c0b86ffb615f794e24525d70", size = 7561528, upload-time = "2026-07-16T18:59:13.288Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/27/1ada0a898c34d009478814fb5a1c2c73cb59036d68563992a1ee0dfcadde/jax_cuda13_plugin-0.11.0-cp314-cp314t-manylinux_2_27_x86_64.whl", hash = "sha256:8f9e14c1b7ef52cccf62594fecd0cde2bb150a0f933e831b0b66b2cb27d91000", size = 7580472, upload-time = "2026-07-16T18:59:14.734Z" },
+]
+
+[package.optional-dependencies]
+with-cuda = [
+    { name = "nvidia-cublas", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cuda-cupti", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cuda-nvcc", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cuda-nvrtc", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cuda-runtime", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cudnn-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cufft", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cusolver", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cusparse", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nccl-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvshmem-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvvm" },
+]
+
 [[package]]
 name = "jaxlib"
 version = "0.11.0"
@@ -425,6 +482,18 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/a1/5a/4d2b1601df3602dba7a14f3348ba9bfe94a18adb428e693df6154c293831/numpy-2.5.1-cp314-cp314t-win_arm64.whl", hash = "sha256:5a6db61f9aaa57e369905c67d852045d3c4f7126405b29d09b19dec118e9c9cb", size = 10697674, upload-time = "2026-07-04T17:07:58.506Z" },
 ]
 
+[[package]]
+name = "nvidia-cublas"
+version = "13.6.1.10"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cuda-nvrtc" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/42/51/a174a0e79793e528ef4645a9d554e3aa48fd8da3f9c9cf523c0176bb121b/nvidia_cublas-13.6.1.10-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:e05a431062a17cb9b02e2f37e67817b637051ce8fad57b388482c594396ddbb4", size = 518610411, upload-time = "2026-07-30T18:18:32.969Z" },
+    { url = "https://files.pythonhosted.org/packages/dd/e9/288a93d8234b8f8ea0ccac7b8cc5d7aa4c663d9676c64f4a2eb3a1f9ffc4/nvidia_cublas-13.6.1.10-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:feb2ed8a1e211bc5774413efc0f1a08c4d5269b56f68b4ac6fe5408e57f7dc1c", size = 410475904, upload-time = "2026-07-30T18:19:26.388Z" },
+]
+
 [[package]]
 name = "nvidia-cublas-cu12"
 version = "12.9.2.10"
@@ -437,6 +506,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/cb/c0/0a517bfe63ccd3b92eb254d264e28fca3c7cab75d07daea315250fb1bf73/nvidia_cublas_cu12-12.9.2.10-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:e4f53a8ca8c5d6e8c492d0d0a3d565ecb59a751b19cfdaa4f6da0ab2104c1702", size = 581240110, upload-time = "2026-04-08T18:52:31.532Z" },
 ]
 
+[[package]]
+name = "nvidia-cuda-cccl"
+version = "13.3.3.4.1"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/96/bd/572971ffc14bd36676c821fc15d991b08fe6179cb09368250147475f954d/nvidia_cuda_cccl-13.3.3.4.1-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:067d19b4b3c9d0f2ebec9f29a311b2863db96bf98e058bbc331597d51ce818cf", size = 3454030, upload-time = "2026-06-29T16:41:49.092Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/ab/049726d90147865a3ea53bae6cb7c35b98bf1fdf96cdb967101329625f83/nvidia_cuda_cccl-13.3.3.4.1-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:cc0adc188d570b09f4d606c7dc05a42aa3d8aa082e0d60f7bbfc5b6435f627c6", size = 3454034, upload-time = "2026-06-29T16:42:07.435Z" },
+]
+
 [[package]]
 name = "nvidia-cuda-cccl-cu12"
 version = "12.9.27"
@@ -446,6 +524,24 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/18/2a/d4cd8506d2044e082f8cd921be57392e6a9b5ccd3ffdf050362430a3d5d5/nvidia_cuda_cccl_cu12-12.9.27-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:37869e17ce2e1ecec6eddf1927cca0f8c34e64fd848d40453df559091e2d7117", size = 3152243, upload-time = "2025-05-01T19:32:13.955Z" },
 ]
 
+[[package]]
+name = "nvidia-cuda-crt"
+version = "13.3.73"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/41/2089e411507d66458d67208bdd1bc562d492bb6458c3d2aea4603072a219/nvidia_cuda_crt-13.3.73-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:60aacc0b5e1e8b40c62abe4d1ab16440add91b99bd2f17f62dd091586b73d166", size = 157353, upload-time = "2026-06-29T16:42:38.163Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/ce/16d76f4b5b3f7460f5ebd17516685495c149c66651ffdd381f90e4d4e65c/nvidia_cuda_crt-13.3.73-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:df14a17ae1c5c3171265411212246654d780f89344ea85344466c6b955247543", size = 157352, upload-time = "2026-06-29T16:43:09.209Z" },
+]
+
+[[package]]
+name = "nvidia-cuda-cupti"
+version = "13.3.75"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d1/7b/d332e362c2c48929f913a5930a596c707c30632ae3248dd16be35f18fc11/nvidia_cuda_cupti-13.3.75-py3-none-manylinux_2_25_aarch64.whl", hash = "sha256:79f5a8c55378a998427d6974b1d85ec91ef684035d016b3fb3b85006914f94f0", size = 11824795, upload-time = "2026-06-29T16:46:00.894Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/68/ee003f64d79bd3cb95e243cfdc906bb0f1ae73b1aa2857baf1d7834dd1ee/nvidia_cuda_cupti-13.3.75-py3-none-manylinux_2_25_x86_64.whl", hash = "sha256:a969cc446c44db612e417bd0af722cfad9009c027dabd4ed255262c0e9a9a3d2", size = 12044533, upload-time = "2026-06-29T16:46:20.662Z" },
+]
+
 [[package]]
 name = "nvidia-cuda-cupti-cu12"
 version = "12.9.79"
@@ -455,6 +551,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/c1/2e/b84e32197e33f39907b455b83395a017e697c07a449a2b15fd07fc1c9981/nvidia_cuda_cupti_cu12-12.9.79-py3-none-manylinux_2_25_x86_64.whl", hash = "sha256:096bcf334f13e1984ba36685ad4c1d6347db214de03dbb6eebb237b41d9d934f", size = 10814997, upload-time = "2025-06-05T20:01:10.168Z" },
 ]
 
+[[package]]
+name = "nvidia-cuda-nvcc"
+version = "13.3.73"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cuda-crt" },
+    { name = "nvidia-cuda-runtime" },
+    { name = "nvidia-nvvm" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5c/14/9f5cdc994d5431e2f08f62ffe34509e7feabd1f2e18517e2d7720c6ff0fd/nvidia_cuda_nvcc-13.3.73-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:70f250825355d2c3aa6c7a972a0ec00f020bad66d2679e527eb4336301c904aa", size = 39515578, upload-time = "2026-06-29T16:47:40.318Z" },
+    { url = "https://files.pythonhosted.org/packages/83/19/e46ef3597ba47a9f8a91ab24533db42a600b659fc418dbe4af0b630bcb41/nvidia_cuda_nvcc-13.3.73-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f483af83166c4fa356a21606076d553b0b4ceaebbd9912e537545080db695bdd", size = 44942138, upload-time = "2026-06-29T16:48:13.615Z" },
+]
+
 [[package]]
 name = "nvidia-cuda-nvcc-cu12"
 version = "12.9.86"
@@ -464,6 +574,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/d6/5c/8cc072436787104bbbcbde1f76ab4a0d89e68f7cebc758dd2ad7913a43d0/nvidia_cuda_nvcc_cu12-12.9.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:44e1eca4d08926193a558d2434b1bf83d57b4d5743e0c431c0c83d51da1df62b", size = 39411138, upload-time = "2025-06-05T20:01:43.182Z" },
 ]
 
+[[package]]
+name = "nvidia-cuda-nvrtc"
+version = "13.3.33"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8b/2c/86916c8a34dcdb0c3ddd1c0e30545041bd781184e437b9cb76fcda70560b/nvidia_cuda_nvrtc-13.3.33-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:82530788b8c6164a54d3fd9ae8bcca8893d397c4aeb998861982a03bbe41e204", size = 51110910, upload-time = "2026-05-26T16:38:16.116Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/b6/60a3641111d39ebfcfcd8b8bfd0290d7623c4b8b5f90952c2d84776f8ca4/nvidia_cuda_nvrtc-13.3.33-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7b05ecda494c6dabc44231a608b060a71008a730d9dfda932cc508e6d29159e0", size = 49260054, upload-time = "2026-05-26T16:37:51.177Z" },
+]
+
 [[package]]
 name = "nvidia-cuda-nvrtc-cu12"
 version = "12.9.86"
@@ -473,6 +592,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/64/eb/c2295044b8f3b3b08860e2f6a912b702fc92568a167259df5dddb78f325e/nvidia_cuda_nvrtc_cu12-12.9.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:096d4de6bda726415dfaf3198d4f5c522b8e70139c97feef5cd2ca6d4cd9cead", size = 44528905, upload-time = "2025-06-05T20:02:29.754Z" },
 ]
 
+[[package]]
+name = "nvidia-cuda-runtime"
+version = "13.3.29"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5f/e5/c1a221c8e6fecd071b80ea44c20fc253ae24f56e15e3f77cfbc3fb76e724/nvidia_cuda_runtime-13.3.29-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:73291e19c9dd919c140c91bda2f80b0eca487da5ee30a086ef7bc4918ecb90ea", size = 2356574, upload-time = "2026-05-26T16:29:56.333Z" },
+    { url = "https://files.pythonhosted.org/packages/97/be/5699b6e642b372f7d24c59c2f41383e2696825e20bab85f7399c7c6a56f7/nvidia_cuda_runtime-13.3.29-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:e04420616e72f563167a7733272992d7e6df6dc5cb54b2f94f9f1520ea9e30c1", size = 2339786, upload-time = "2026-05-26T16:30:21.584Z" },
+]
+
 [[package]]
 name = "nvidia-cuda-runtime-cu12"
 version = "12.9.79"
@@ -494,6 +622,30 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/10/13/b8887c869cf2471339a24b60d3c28e761facbb534935f572b61423371abb/nvidia_cudnn_cu12-9.24.0.43-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:f424192dd85e7d29f44be18df2dae4c80d32c67a29c0d42f5c283c40cfdf871c", size = 799083985, upload-time = "2026-07-02T16:25:37.467Z" },
 ]
 
+[[package]]
+name = "nvidia-cudnn-cu13"
+version = "9.24.0.43"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cublas" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ca/30/7c257e3d5cb4fecb147b93895c66e29c93f8e76d74b45bb418ff0587c4ec/nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:a6812a554a1ff0413e9c52b84c26c050380649ab9615f9c16bded368ce9f421f", size = 650976863, upload-time = "2026-07-02T16:23:39.248Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/ba/791cffd048fe5b044e620df55267e3e95c0e6e07d50b41e377c03dfc910f/nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:71f181cd810e90f9b6023b01186fe82d13d65f0ec098581ee201d39fad769e4b", size = 553099438, upload-time = "2026-07-02T16:27:42.58Z" },
+]
+
+[[package]]
+name = "nvidia-cufft"
+version = "12.3.0.29"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-nvjitlink" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f9/bc/543e6dfbcb659ed07a94a72d8925c8b5688dabc1f616ab2f6575483a3cb6/nvidia_cufft-12.3.0.29-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:3c268ec73fdbfd22e42913dcef6e49e0c305ab09cd9848268a841cbe9f33f748", size = 184700143, upload-time = "2026-05-26T16:45:25.443Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/00/fab4a29fa1d7eb43bc6b94de4e86312c5e425d5582e58b9641300b9dffc7/nvidia_cufft-12.3.0.29-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:edb25c0626bd202ee5acc035b5dd361a3b89ed3b75a81a52df72c89150cb57c2", size = 184730406, upload-time = "2026-05-26T16:46:18.193Z" },
+]
+
 [[package]]
 name = "nvidia-cufft-cu12"
 version = "11.4.1.4"
@@ -506,6 +658,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/95/f4/61e6996dd20481ee834f57a8e9dca28b1869366a135e0d42e2aa8493bdd4/nvidia_cufft_cu12-11.4.1.4-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:c67884f2a7d276b4b80eb56a79322a95df592ae5e765cf1243693365ccab4e28", size = 200877592, upload-time = "2025-06-05T20:05:45.862Z" },
 ]
 
+[[package]]
+name = "nvidia-cusolver"
+version = "12.2.6.9"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cublas" },
+    { name = "nvidia-cusparse" },
+    { name = "nvidia-nvjitlink" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/13/bd/61dcd7917f64c9e81a35f939f1bf4ab6535adf712192e915fb472e5bbf6c/nvidia_cusolver-12.2.6.9-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:658ce9e6ed8a321033419fe6e55faf0f245a15aa23669eda03d7f2dcce436d73", size = 267275679, upload-time = "2026-06-29T16:59:54.924Z" },
+    { url = "https://files.pythonhosted.org/packages/1c/3a/23e46a0919ba5ecf1a864400542a6c78d18c76893a0a77e1856e7033af31/nvidia_cusolver-12.2.6.9-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:e32cb40c8a4d3df475b556caaf6af5808ef064df7e291a049528a0d28017b674", size = 244667799, upload-time = "2026-06-29T17:00:42.543Z" },
+]
+
 [[package]]
 name = "nvidia-cusolver-cu12"
 version = "11.7.5.82"
@@ -520,6 +686,18 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/33/40/79b0c64d44d6c166c0964ec1d803d067f4a145cca23e23925fd351d0e642/nvidia_cusolver_cu12-11.7.5.82-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:15da72d1340d29b5b3cf3fd100e3cd53421dde36002eda6ed93811af63c40d88", size = 338117415, upload-time = "2025-06-05T20:07:16.809Z" },
 ]
 
+[[package]]
+name = "nvidia-cusparse"
+version = "12.8.2.51"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-nvjitlink" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ac/bc/925058f9343d93426864cbb113a3f5cf6db97a7d11642aaf4159b6f0c57a/nvidia_cusparse-12.8.2.51-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:00469fcf62c4d464a1225abd9b20864ecff35e3fbc9fb992572e83d358927755", size = 172868683, upload-time = "2026-06-29T17:01:23.244Z" },
+    { url = "https://files.pythonhosted.org/packages/a6/dc/752e401a5d6fc5f1948cf190add5afc27e440e5ca08bc3fab66826c16213/nvidia_cusparse-12.8.2.51-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:65cbcc4e37a34fca4ee7df2fd57da103593842cda1bbb4a144664ecfe59873a5", size = 154538196, upload-time = "2026-06-29T17:02:06.078Z" },
+]
+
 [[package]]
 name = "nvidia-cusparse-cu12"
 version = "12.5.10.65"
@@ -541,6 +719,24 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/50/32/e7ffa9c324ae260e5dbb4af2cd557bf7a8d155c8ac7b79a785fe1796fb92/nvidia_nccl_cu12-2.30.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:8ce1b8213f61f2bfac132e6df890af6450b77cbd140c6ce4e98cb0c2d8e678c9", size = 303361239, upload-time = "2026-06-09T03:24:53.816Z" },
 ]
 
+[[package]]
+name = "nvidia-nccl-cu13"
+version = "2.30.7"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d1/21/a73174c6157101bdf1ffc22b517f76ff0082613989dd9bc8f43e8034caac/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:ca786ffa5a647c75d4d1f5cc72a6c4f537947e2ba8823d7c8aaf768e7a7b9f77", size = 215983881, upload-time = "2026-06-09T03:23:15.633Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/34/c500f90c7ae641b8e0f98965b36b8a7ac79cc8b296e8d251fe3eb592ee54/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:cefa7fdb9710efd0f39c5f1be1d61ff6fc9a996c451265bd7fbdcf9455ed4b50", size = 215965170, upload-time = "2026-06-09T03:23:39.73Z" },
+]
+
+[[package]]
+name = "nvidia-nvjitlink"
+version = "13.3.33"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f0/ee/580ca6f29dcab0221db8706badca1bbbb084f1975c4d4e83329c3a7e31f0/nvidia_nvjitlink-13.3.33-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:26a6de7fb4c8fdaa7703d3dad720d6d427ddfea5c48a528fd97c11733ad830e5", size = 40742423, upload-time = "2026-05-26T16:54:51.613Z" },
+    { url = "https://files.pythonhosted.org/packages/69/30/45414e35ff2eee7db3da037e5707037ccf9d2b5218ffbdb055ea4d5aa98a/nvidia_nvjitlink-13.3.33-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:ce48b37dfeb3cb1eae4cf85adacb47d7a6539ea2272870c9a3628ce275c2037e", size = 39168635, upload-time = "2026-05-26T16:54:13.906Z" },
+]
+
 [[package]]
 name = "nvidia-nvjitlink-cu12"
 version = "12.9.86"
@@ -562,6 +758,28 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/63/03/cde130e3ff706784f9efd47204f299b15113ea08203e98a556603dac245d/nvidia_nvshmem_cu12-3.7.2-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:02b6d3482d90ea8ba214bc400362297f573257763990a98c9846ec5c786a7227", size = 230171132, upload-time = "2026-07-17T19:25:07.092Z" },
 ]
 
+[[package]]
+name = "nvidia-nvshmem-cu13"
+version = "3.7.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cuda-cccl" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/81/11/1767ef120299972822a43aceffbc81ac813aad2089d817e36e4e12aa4c9c/nvidia_nvshmem_cu13-3.7.2-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:0e748308b9e0a8a8d81859353f3e6c98fbe85a85f652a915a2ba60149a0bd124", size = 135174619, upload-time = "2026-07-17T19:24:15.404Z" },
+    { url = "https://files.pythonhosted.org/packages/4f/4a/037dfbbafc96ad1828ef62326f27970cae2c0eb271cf5b8e6922f32a7427/nvidia_nvshmem_cu13-3.7.2-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:b9a4f843c62b5dded8ad33fd6d220e007dcd39f06cee549292f5ba14ce4c235c", size = 135397797, upload-time = "2026-07-17T19:25:26.4Z" },
+]
+
+[[package]]
+name = "nvidia-nvvm"
+version = "13.3.73"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f3/e7/ff646aa6015c7e6d12aad234e68925c87b6681d8d18c3ac40535994a3b0d/nvidia_nvvm-13.3.73-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:0e28e0858a3475e11ac67d35301cd5bf82666a1c0dc4ec4e80ceaf3a5fd1dea8", size = 69250424, upload-time = "2026-06-29T17:08:07.453Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/05/35754a7105563fd9b496e5ee8e1acd986aef8258760c3cbccf419aee861a/nvidia_nvvm-13.3.73-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e2bcdd5783b5481445f1f0e7170cb836cc0d72999839ba850bbba6dc97b76bb8", size = 66984478, upload-time = "2026-06-29T17:07:43.765Z" },
+    { url = "https://files.pythonhosted.org/packages/ad/6b/d5756f485012b920475cbc01457c1b9a7d0485bfb04b92598c5e1ef3e9ab/nvidia_nvvm-13.3.73-py3-none-win_amd64.whl", hash = "sha256:b5c91dfa59ee4cee90b2dfb19c6203f31c914b9c9b5ca10726c2da7cf8ed401d", size = 59981103, upload-time = "2026-06-29T17:21:43.334Z" },
+]
+
 [[package]]
 name = "omegaconf"
 version = "2.3.1"