Availability¶
Not every checkpoint has every value. A model loaded with sdpa never builds an attention pattern, and OPT has no mlp module. model.support() says which values this checkpoint has before any trace, from the config and the module tree alone. None means available, and anything else is the reason, a string. A family's page in the encyclopedia shows the same report for each of its public checkpoints, with the reason beside every value the checkpoint lacks.
from nnterp import StandardizedTransformer
model = StandardizedTransformer("openai-community/gpt2", dispatch=True, device="cpu")
support = model.support()
print(support["layer_output"])
print(support["self_attn.attention_output"])
print(support["self_attn.attention_probabilities"][0])
print(len(support["self_attn.attention_probabilities"]))
# None
# None
# read inside the eager attention forward, but this model runs 'sdpa'; load with attn_implementation='eager'
# 12
Loaded with attn_implementation="eager", every entry on GPT-2 is None.
Three forms¶
model.support() is the root's values plus every block value, by dotted name. A block value is None when every block has it, else {layer: reason} for the blocks that lack it, which is why the entry above is a dict with 12 keys. model.support(layer=i) is one block, flat, and envoy.support() is one envoy's own values without the prefix.
print(model.support(layer=0)["self_attn.attention_scores"])
print(model.layers[0].mlp.support())
# read inside the eager attention forward, but this model runs 'sdpa'; load with attn_implementation='eager'
# {'mlp_activation': None, 'mlp_neurons': None, 'mlp_output': None}
The keys come from the tree. Every child of a block that carries standard values is listed under its standard name, so a value you add through envoys= appears as self_attn.<name> or mlp.<name>. A module that some blocks lack is reported on those blocks, and a module no block has (OPT's mlp) has no key at all, so guard with support.get("mlp.mlp_output") on a checkpoint that may lack it. A hybrid reads as a per-block dict. GPT-2 has every module, which is why this block uses Qwen3.5-0.8B, where each block carries either linear_attn or self_attn.
hybrid = StandardizedTransformer("Qwen/Qwen3.5-0.8B", attn_implementation="eager")
support = hybrid.support()
print(sorted(support["self_attn.attention_output"]))
print(support["self_attn.attention_output"][0])
print(sorted(support["linear_attn.attention_output"]))
print([i for i, layer in enumerate(hybrid.layers) if getattr(layer, "self_attn", None) is not None])
# [0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20, 21, 22]
# no self_attn module on this block
# [3, 7, 11, 15, 19, 23]
# [3, 7, 11, 15, 19, 23]
No dispatch=True here: support() reads the config and the tree, which a meta-device model has.
The reasons¶
| reason, as it starts | values | when |
|---|---|---|
read inside the eager attention forward, but this model runs 'sdpa' |
the attention interior | loaded with any attn_implementation but "eager" |
this checkpoint sets reorder_and_upcast_attn |
the GPT-2 interior | a GPT-2 config with that flag, which takes GPT-2's own attention path |
no self_attn module on this block, no linear_attn module on this block |
every value of that module | a hybrid's blocks: some other block has the module and this one does not |
a parallel block: the attention and the MLP read the same input, so there is no stream between them |
layer_mid |
a parallel block (Cohere, GPT-NeoX with use_parallel_residual): the contribution identity still holds, and only layer_mid has no tensor |
no act_fn module on this mlp, no down_proj module on this mlp |
mlp_activation, mlp_neurons |
the family aliases no module to that name, such as an activation applied as a function |
routed experts; read per-expert via the Moe component |
mlp_activation, mlp_neurons |
the block's MLP is a mixture of experts (Mixture of experts) |
the state after each token is materialized only by the token-by-token kernel |
linear_attn.state, linear_attn.states |
a gated DeltaNet mixer before nnterp.route_kernels(model.family, "torch") (Gated DeltaNet) |
the default family cannot tell what this sublayer adds to the stream |
attention_output, mlp_output |
a model_type with no family (Loading) |
The reasons are evaluated on the instance, so a change after the load changes the answer: route_kernels turns the state entries to None. support() predicts from the config; if a run's forward then takes a path the family does not expect, the read raises nnsight's SourceNotAvailable naming the missing operation (Intermediate Operations).
nnterp.Unavailable at the read¶
Reading or writing an unavailable value raises nnterp.Unavailable with the same reason, at that line, before the model runs.
import nnterp
try:
with model.trace("The Eiffel Tower is in"):
pattern = model.layers[0].self_attn.attention_probabilities.save()
except nnterp.Unavailable as error:
print(error)
# model.transformer.h.0.attn.attention_probabilities is not available: read inside the eager attention forward, but this model runs 'sdpa'; load with attn_implementation='eager'
A module that does not exist on a block is an ordinary missing attribute: model.layers[0].mlp on OPT raises AttributeError, and getattr(layer, "mlp", None) is None, which is the form to pick blocks with outside the trace.
hasattr¶
hasattr treats only AttributeError as absence. An unavailable value raises Unavailable through it, and an available value read outside a trace raises nnsight's ValueError, so hasattr never answers False. Check support() outside the trace instead.
try:
hasattr(model.layers[0].self_attn, "attention_probabilities")
except nnterp.Unavailable as error:
print(type(error).__name__)
try:
hasattr(model.layers[0], "layer_output")
except ValueError as error:
print(error)
# Unavailable
# Cannot access `model.transformer.h.0.output` outside of interleaving
Guarding code¶
Branch on support() outside the trace, and the body reads only what exists.
model = StandardizedTransformer("openai-community/gpt2", dispatch=True, device="cpu", attn_implementation="eager")
support = model.support()
has_pattern = support["self_attn.attention_probabilities"] is None
has_mlp = support.get("mlp.mlp_output", "absent") is None
with model.trace("The Eiffel Tower is in"):
if has_pattern:
pattern = model.layers[0].self_attn.attention_probabilities.save()
if has_mlp:
mlp = model.layers[0].mlp.mlp_output.save()
print(pattern.shape, mlp.shape)
# torch.Size([1, 12, 7, 7]) torch.Size([1, 7, 768])
The repr lists a value whether or not this checkpoint has it. Only a value a family marks unavailable in its class body prints as Unavailable: <reason>, and a config-dependent reason shows in support() alone. Once a value is available, its shape is on Layouts.