Hotfix/inference

docs/architectures.md

@@ -32,7 +32,7 @@ All stateless architectures apply the checkpoint's text-only `config.max_len`
 before encoding labels. Whether labels share a later backbone context with the
 text depends on the architecture: uni-encoders share it, while bi-encoders keep
 the two sequences separate. Cached StreamingSpan sessions use their causal
-decoder context instead. See [Input limits and truncation](input_limits.md) for
+decoder context instead. See [Input limits and truncation](usage.md#input-limits-and-truncation) for
 the complete behavior and production preflight guidance.
 
 ## Vanilla GLiNER (UniEncoderSpan)
@@ -660,7 +660,7 @@ complete current entity snapshot with document-relative character offsets. It
 is not a delta: callers that render live changes should compare consecutive
 snapshots to identify additions, updates, and removals.
 
-The [StreamingSpan guide](streaming.md) covers the session API, persistent
+The [StreamingSpan guide](usage.md#streamingspan-models) covers the session API, persistent
 batches, asynchronous microbatching, cache cleanup, configuration, and
 deployment trade-offs. A PII-oriented checkpoint is available at
 [knowledgator/gliner-stream-pii-v1.0](https://huggingface.co/knowledgator/gliner-stream-pii-v1.0).

docs/configs.md

@@ -163,7 +163,7 @@ configurations may override the base default.
 This is not necessarily the backbone's total context length. Uni-encoder
 prompts and text share the later transformer subword/context limit, while
 bi-encoders encode labels separately. See [Input limits and
-truncation](input_limits.md) for all limits and the StreamingSpan exception.
+truncation](usage.md#input-limits-and-truncation) for all limits and the StreamingSpan exception.
 
 ---
 

docs/convert_to_onnx.md

@@ -16,16 +16,16 @@ automatically; keep an OpenVINO `.xml` file beside its matching `.bin` file.
 
 ## Installation
 
-ONNX Runtime CPU support is included in the default installation:
+For ONNX Runtime CPU support, install the optional `onnx` extra:
 
 ```bash
-pip install gliner
+pip install "gliner[onnx]"
 ```
 
 Install the Python `onnx` package when exporting a PyTorch checkpoint to ONNX:
 
 ```bash
-pip install gliner onnx
+pip install "gliner[onnx]" onnx
 ```
 
 For ONNX Runtime with CUDA execution providers, install the GPU extra:
@@ -35,6 +35,9 @@ pip install "gliner[gpu]"
 ```
 
 Add `onnx` to the same command if this environment will also export models.
+Choose only one of `gliner[onnx]` and `gliner[gpu]`: both runtime packages
+provide the same Python module and must not be installed together. The default
+`gliner` installation does not install either ONNX Runtime package.
 
 For direct OpenVINO export and inference, install the OpenVINO extra. Direct
 OpenVINO conversion does not create an intermediate ONNX model:

docs/index.md

@@ -10,9 +10,7 @@ This documentation includes installation guides, tutorials, advanced topics, and
 intro
 instalation
 quickstart
-input_limits
 usage
-streaming
 configs
 training
 architectures

docs/instalation.md

@@ -9,12 +9,15 @@ To begin using the GLiNER model, you can install the GLiNER Python library throu
 pip install gliner
 ```
 :::tip[ONNX Runtime]
-CPU inference with ONNX Runtime is included in the default installation. Install the separate `onnx` package to export models, or use the GPU feature for CUDA execution providers.
+ONNX Runtime is optional. Choose the `onnx` extra for CPU inference or the `gpu` extra for CUDA execution providers. Install only one: the CPU and GPU packages share the same Python module. Install the separate `onnx` package to export models.
 :::
 
 ```bash
-pip install onnx             # required for ONNX export
+pip install "gliner[onnx]"  # CPU-backed ONNX Runtime
+# or
 pip install "gliner[gpu]"   # GPU-backed ONNX Runtime
+
+pip install onnx           # additionally required for ONNX export
 ```
 
 :::tip[OpenVINO]

docs/intro.md

@@ -85,13 +85,16 @@ To begin using the GLiNER model, you can install the GLiNER Python library throu
 !pip install gliner
 ```
 
-CPU ONNX Runtime support is included in the default installation. Install
-`onnx` when exporting models, or install GLiNER with the GPU feature for CUDA
-execution providers:
+ONNX Runtime is optional. Choose the `onnx` extra for CPU inference or the
+`gpu` extra for CUDA execution providers. Install only one: the CPU and GPU
+packages share the same Python module. Also install `onnx` when exporting models:
 
 ```bash
-!pip install onnx
-!pip install "gliner[gpu]"
+!pip install "gliner[onnx]"  # CPU
+# or
+!pip install "gliner[gpu]"   # NVIDIA GPU
+
+!pip install onnx           # additionally required for ONNX export
 ```
 
 Install OpenVINO support to export models directly to OpenVINO IR or run them

docs/serving.md

@@ -80,7 +80,11 @@ outputs = client.predict(
 ```
 
 Each dictionary key is returned as the entity label; its value is the prompt encoded
-for that text. A single dictionary is shared across every input text.
+for that text. A single dictionary is shared across every input text. When supplying
+a list of dictionaries, its length and order must match the input texts, and each
+dictionary contains all labels for its corresponding text. For a single text, put
+all labels in one dictionary. See [Label descriptions](usage.md#label-descriptions)
+for examples of shared and per-text label sets.
 
 Network or server errors surface as `gliner.serve.client.GLiNERClientError`.
 
@@ -90,7 +94,7 @@ overrides the loaded model's text-only `config.max_len`; requests above it can
 still return HTTP success after GLiNER keeps only the prefix and emits a Python
 warning inside the replica. If the service contract requires full-input
 coverage, validate or window requests explicitly. See [Input limits and
-truncation](input_limits.md) for a preflight that uses the public model API.
+truncation](usage.md#input-limits-and-truncation) for a preflight that uses the public model API.
 :::
 
 **HTTP request (no client library):**