llmdbg

llmdbg is an early-stage debugger for stepping through language-model layers and presenting human-readable interpretations of internal vectors. It speaks the Debug Adapter Protocol so editor integrations can use a standard debugger UI.

The current prototype includes real GPT-2 Small loading, layer-state capture, logit-lens interpretations, and a VS Code debug extension. A deterministic built-in trace remains available when no model is configured.

pip install --pre llmdbg
llmdbg --version

VS Code prototype

Build and install the extension:

cd vscode-extension
npm install
npm run package
code --install-extension llmdbg-0.0.1.vsix

The recommended format is now a real Python program named gpt2.llmdbg.py:

from llmdbg import debug, gpt2

model = gpt2("openai-community/gpt2", device="cuda", dtype="bfloat16")
prompt = "The capital of France is"
token = 3
tokens = model.tokenize(prompt)

def attention(x, layer):
    normalized = layer.layernorm(x)
    qkv = layer.qkv_projection(normalized)
    mixed = layer.attention_mixing(qkv)
    update = layer.output_projection(mixed)
    return x + update

def mlp(x, layer):
    normalized = layer.layernorm(x)
    hidden = layer.expansion(normalized)
    hidden = layer.gelu(hidden)
    update = layer.projection(hidden)
    return x + update

with debug(model, tokens, token=token):
    x = model.token_embedding(tokens)
    for layer in model.layers:
        x = attention(x, layer)
        x = mlp(x, layer)
    x = model.final_layernorm(x)
    logits = model.vocabulary_projection(x)

This is executable Python. Model and tensor objects are handles over llmdbg's runtime; hidden[1843] = 0 records an activation intervention and recomputes the selected token. Python indexing and slicing such as hidden[:10] work in source and Watch. The original compact gpt2.llmdbg format remains supported:

model = "openai-community/gpt2"
revision = "main"
device = "cuda"
dtype = "bfloat16"
prompt = "The capital of France is"
token = 3
stop_on_entry = false

def attention():
    layernorm
    qkv_projection
    attention_mixing
    output_projection
    residual_add

def mlp():
    layernorm
    expansion
    gelu
    projection
    residual_add

submodule = "token_embedding"

layer = 0
attention()
mlp()

layer = 1
attention()
mlp()

Open either file and press F5. If VS Code asks for a debugger, choose llmdbg. No launch.json is required. If a workspace launch configuration is preferred, the complete minimal version is:

{
  "version": "0.2.0",
  "configurations": [
    {
      "type": "llmdbg",
      "request": "launch",
      "name": "llmdbg: Active file",
      "config": "${file}"
    }
  ]
}

On first use, the extension creates a private Python environment in VS Code's extension storage. It does not install anything into the workspace. The first GPT-2 run downloads and loads the model, then keeps its tensors resident in a persistent model-host process. Stopping and restarting the debugger reuses the resident model and runs only a new forward pass. Closing/reloading VS Code stops the host and releases the model. After layer = N, attention() and mlp() resolve to that transformer block and become authored execution stops. Set breakpoints on these calls or use Step Over to advance through them. Runtime vectors appear in debugger scopes; the source file is never rewritten.

Press Step Into (F11) on attention() or mlp() to enter its authored definition in the same file. Attention exposes LayerNorm, the combined QKV projection, the concatenated per-head attention result, output projection, and residual add. MLP exposes LayerNorm, expansion, GELU, projection, and residual add. F10 advances inside a definition; Shift+F11 returns to the next model operation. The call site remains visible as the parent stack frame.

token = 3 makes token position 3 the default for every ordinary source breakpoint. To override it for one breakpoint, use the shorter conditional expression token=4. Positions are zero-based. The debugger prefills tokens before the selected position into the KV cache, then captures that token with a one-token forward pass. Stepping beyond the last selected layer advances to the next token using the updated cache.

Watch one zero-based dimension with residual_stream[42], input[42], output[42], or delta[42]. The Residual stream (output) debugger scope exposes the current hidden representation explicitly. Run llmdbg: Plot Current Vector from the Command Palette or debug toolbar to inspect the full vector as a dimension plot, histogram, or top-activation chart. An open plot updates automatically after each debugger stop.

Attention stops expose the 768D LayerNorm output, combined 2304D QKV projection, 768D concatenated attention-head result, and 768D output projection. MLP stops expose the 768D LayerNorm output, the 3072D linear expansion before activation, the 3072D expansion after GELU, and the 768D projection. Use indexed Watches such as layernorm[42], mlp_expansion_pre_activation[42], mlp_expansion[42], and mlp_projection[42]; all available vectors also appear in the plot selector.

Expanding Top activated neurons at an MLP stop lazily ranks the five largest post-GELU activations. Each neuron shows its output-direction norm, the top five tokens from W_down[neuron] @ W_unembed, and the top five direct token contributions after scaling by the current activation. Token IDs and decoded text are both displayed. Results are cached for the current token and layer. Add top_neurons to Watch to keep this tree visible while stepping; VS Code reevaluates it at every stop and preserves the expandable neuron/token structure. At non-MLP stops it reads available at MLP stops.

The yellow execution arrow identifies the next submodule to execute. Step Over executes that submodule, moves to the next statement, and exposes the completed operation under Previous output and its detail scopes. Watch dimensions are writable through VS Code's Set Value action. For example, set output[42], mlp_expansion[1843], or top_neurons[0].activation; llmdbg records the intervention, replays the selected token from its KV-cache prefix, and refreshes all downstream values. Rank-based top_neurons[0] may identify a different neuron after mutation; use mlp_expansion[neuron_id] when targeting a fixed neuron.

License

MIT