fimmy-small-java

Java-specialized code completion model.

Architecture

GPT-2 variant with language-conditional LayerNorm (12 layers, 12 heads, dim=768, 126M params).

• Per-language affine after LayerNorm: x * (1 + scale[lang_id]) + shift[lang_id]
• KV cache for efficient autoregressive generation
• Beam search + temperature/top-k/top-p sampling
• Context extension (YaRN/periodic/linear) from 1024 to 16k+
• wne suffix positional embedding for FIM
• GPT-2 BPE tokenizer (vocab_size=50257)

Quick Start

from modeling_fimmy import load_fimmy_model
from transformers import GPT2Tokenizer
import torch

model = load_fimmy_model("di-zhang-fdu/fimmy-small-java")
tok = GPT2Tokenizer.from_pretrained("gpt2")
LANG_ID = 2

# Greedy completion
ids = tok.encode("def hello():
    ", return_tensors="pt")
out = model.generate(ids, lang_id=LANG_ID, max_new_tokens=15)
print(tok.decode(out[0, ids.shape[1]:]))

# Beam search
out = model.generate(ids, lang_id=LANG_ID, beam_width=3, max_new_tokens=15)

# Temperature sampling
out = model.generate(ids, lang_id=LANG_ID, temperature=0.7, top_k=5, top_p=0.9)

# Extend to 16k context (zero quality loss for original 1024)
model.transformer.extend_positional_embeddings(scale=16, method="yarn")

Fill-In-the-Middle (FIM)

FIM completes code between a before (prefix) and after (suffix) context.

before = "def add(a, b):
    return "
after = "a + b

result = add(1, 2)"

tok = GPT2Tokenizer.from_pretrained("gpt2")

# Method 1: Suffix-first FIM (recommended)
# Put suffix before prefix so the model sees both contexts via causal attention
after_ids = tok.encode(after)
before_ids = tok.encode(before)
combined = torch.tensor([after_ids + before_ids])
out = model.generate(combined, lang_id=LANG_ID, max_new_tokens=10)
print(tok.decode(out[0, combined.shape[1]:]))

# Method 2: Prefix-only completion (no suffix context)
ids = tok.encode(before, return_tensors="pt")
out = model.generate(ids, lang_id=LANG_ID, max_new_tokens=10)
print(tok.decode(out[0, ids.shape[1]:]))

# Method 3: wne-based FIM (experimental)
# Uses wne (suffix positional embedding) to encode suffix position
def fim_forward(model, before_ids, after_ids, lang_id):
    all_ids = torch.cat([before_ids, after_ids], dim=1)
    B, T = all_ids.shape
    with torch.no_grad():
        pos = torch.arange(T).unsqueeze(0)
        x = model.transformer.wte(all_ids) + model.transformer.wpe(pos)
        # Add wne positional embedding to suffix tokens
        after_len = after_ids.shape[1]
        if after_len > 0:
            suffix_pos = torch.arange(after_len).unsqueeze(0)
            x[0, before_ids.shape[1]:] += model.transformer.wne(suffix_pos)[0]
        for block in model.transformer.h:
            x, _ = block(x, lang_id=lang_id)
        x = model.transformer.ln_f(x, lang_id)
        logits = model.lm_head(x)
    return logits[0, before_ids.shape[1] - 1]

next_logits = fim_forward(model, before_ids, after_ids, lang_id=LANG_ID)
print(tok.decode([next_logits.argmax().item()]))

Configuration

Field	Value
n_layer	12
n_head	12
n_embd	768
intermediate_size	3072
vocab_size	50257
n_positions	1024 (extendable to 16k+)
n_lang	30
language	java
language_id	2

All Models

• fimmy-nano — 6L, 83M, generic
• fimmy-small-python — 12L, 126M, python
• fimmy-small-java — 12L, 126M, java (this model)
• fimmy-small-go — 12L, 126M, go
• fimmy-small-ruby — 12L, 126M, ruby
• fimmy-small-rust — 12L, 126M, rust
• fimmy-small-cpp — 12L, 126M, cpp
• fimmy-small-dart — 12L, 126M, dart
• fimmy-small-julia — 12L, 126M, julia
• fimmy-small-hcl — 12L, 126M, hcl
• fimmy-small-generic — 12L, 126M, generic
• fimmy-medium-generic — 24L, 359M, generic

License

MIT