Expand API reference into per-symbol PyTorch-style pages

docs/api/compiler/TvarantTracer.md

@@ -0,0 +1,18 @@
+# TvarantTracer
+
+```python
+class torch_tvarant.compiler.TvarantTracer
+```
+
+::: torch_tvarant.compiler.TvarantTracer
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/compile.md

@@ -0,0 +1,18 @@
+# compile
+
+```python
+torch_tvarant.compiler.compile(module, *, dynamic=False) → Callable
+```
+
+::: torch_tvarant.compiler.compile
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/compile_fx.md

@@ -0,0 +1,18 @@
+# compile_fx
+
+```python
+torch_tvarant.compiler.compile_fx(gm, example_inputs=None) → GraphModule
+```
+
+::: torch_tvarant.compiler.compile_fx
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/fuse_gemm_epilogue.md

@@ -0,0 +1,18 @@
+# fuse_gemm_epilogue
+
+```python
+torch_tvarant.compiler.fuse_gemm_epilogue(gm) → GraphModule
+```
+
+::: torch_tvarant.compiler.fuse_gemm_epilogue
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/fuse_pointwise.md

@@ -0,0 +1,18 @@
+# fuse_pointwise
+
+```python
+torch_tvarant.compiler.fuse_pointwise(gm) → GraphModule
+```
+
+::: torch_tvarant.compiler.fuse_pointwise
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/index.md

@@ -0,0 +1,20 @@
+# torch_tvarant.compiler
+
+Graph compiler that fuses GEMM epilogues and pointwise chains for inference on
+`device="tvarant"`.
+
+```python
+import torch_tvarant
+compiled = torch_tvarant.compiler.compile(model)
+y = compiled(x)
+```
+
+| API | Summary |
+|---|---|
+| [`compile`](compile.md) | Compile a module for inference |
+| [`compile_fx`](compile_fx.md) | Fuse an FX graph |
+| [`trace_module`](trace_module.md) | FX-trace with inlined Linear/activations |
+| [`fuse_gemm_epilogue`](fuse_gemm_epilogue.md) | Fuse mm/linear + act into `linear_act` |
+| [`fuse_pointwise`](fuse_pointwise.md) | Collapse pointwise chains |
+| [`register`](register.md) | Register Dynamo backend `"tvarant"` |
+| [`TvarantTracer`](TvarantTracer.md) | Custom FX tracer |

docs/api/compiler/register.md

@@ -0,0 +1,18 @@
+# register
+
+```python
+torch_tvarant.compiler.register() → None
+```
+
+::: torch_tvarant.compiler.register
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/compiler/trace_module.md

@@ -0,0 +1,18 @@
+# trace_module
+
+```python
+torch_tvarant.compiler.trace_module(module) → GraphModule
+```
+
+::: torch_tvarant.compiler.trace_module
+    options:
+      show_root_heading: false
+      show_root_toc_entry: false
+      show_source: true
+      show_signature: true
+      show_signature_annotations: true
+      separate_signature: true
+      merge_init_into_class: true
+      members_order: source
+      docstring_section_style: table
+      heading_level: 2

docs/api/index.md

@@ -1,12 +1,24 @@
 # API reference
 
-This section mirrors a PyTorch-style API layout for **torch_tvarant**.
+PyTorch-style reference for **torch_tvarant**. Each entry is a dedicated page
+with signature, parameters, and examples.
 
-| Page | Contents |
-|---|---|
-| [torch.tvarant](torch.md) | Device module helpers + Python package / compiler (auto-generated) |
-| [Supported ops](ops.md) | ATen and custom ops registered for `PrivateUse1` |
-| [Kernels](kernels.md) | Host and OpenCL kernel catalog |
-| [C++](cpp.md) | Extension layout, `TORCH_LIBRARY_IMPL`, adding ops |
+## torch.tvarant
 
-Unimplemented ATen ops fall back to CPU; see [Supported ops](ops.md).
+Device module — [`torch.tvarant` overview](torch/index.md)
+
+- [`is_available`](torch/is_available.md) · [`device_count`](torch/device_count.md) · [`backend`](torch/backend.md)
+- [`set_device`](torch/set_device.md) · [`synchronize`](torch/synchronize.md)
+- [`manual_seed`](torch/manual_seed.md) · [`get_rng_state`](torch/get_rng_state.md) · …
+
+## torch_tvarant.compiler
+
+- [`compile`](compiler/compile.md) · [`compile_fx`](compiler/compile_fx.md)
+- [`trace_module`](compiler/trace_module.md) · [`register`](compiler/register.md)
+- [Compiler overview](compiler/index.md)
+
+## Supported ops
+
+- [Ops index](ops/index.md) — `relu`, `add`, `mm`, `linear`, `linear_act`, …
+- [Kernels](kernels.md) — host / OpenCL catalog
+- [C++](cpp.md) — extension layout & `TORCH_LIBRARY_IMPL`

docs/api/ops/add.md

@@ -0,0 +1,33 @@
+# add
+
+```python
+torch.add(input, other, *, alpha=1, out=None) → Tensor
+input + other
+```
+
+Adds `other`, scaled by `alpha`, to `input`:
+
+$$
+\text{out}_i = \text{input}_i + \alpha \times \text{other}_i
+$$
+
+## Parameters
+
+- **input** (`Tensor`) – First operand (Tvarant).
+- **other** (`Tensor or Number`) – Second operand (broadcasting).
+- **alpha** (`Number`) – Multiplier for `other`. Default: `1`.
+
+## Shape
+
+- Input: $(*)$
+- Output: broadcast shape of `input` and `other`
+
+## Example
+
+```python
+y = torch.add(a, b, alpha=1.0)  # or a + b
+```
+
+---
+
+[← All supported ops](index.md)

docs/api/ops/clone.md

@@ -0,0 +1,27 @@
+# clone / contiguous
+
+```python
+Tensor.clone(memory_format=torch.preserve_format) → Tensor
+Tensor.contiguous(memory_format=torch.contiguous_format) → Tensor
+```
+
+Materialize a copy (`clone`) or a contiguous storage (`contiguous`) on Tvarant via `_copy_from`.
+
+## Parameters
+
+- **memory_format** (`torch.memory_format`) – Desired memory format.
+
+## Shape
+
+- Output: same logical shape as input
+
+## Example
+
+```python
+y = x.clone()
+c = x.t().contiguous()
+```
+
+---
+
+[← All supported ops](index.md)

docs/api/ops/embedding.md

@@ -0,0 +1,27 @@
+# embedding
+
+```python
+torch.nn.functional.embedding(input, weight, ...) → Tensor
+```
+
+Lookup embeddings. Indices should be convertible to int32 on the Tvarant path.
+
+## Parameters
+
+- **input** (`LongTensor`) – Token indices.
+- **weight** (`Tensor`) – Embedding table `(V, D)` on Tvarant.
+
+## Shape
+
+- Input: $(*)$
+- Output: $(*, D)$
+
+## Example
+
+```python
+e = torch.nn.functional.embedding(tokens, weight)
+```
+
+---
+
+[← All supported ops](index.md)

docs/api/ops/empty.md

@@ -0,0 +1,28 @@
+# empty
+
+```python
+torch.empty(*size, *, dtype=None, layout=torch.strided, device=None, pin_memory=False, memory_format=torch.contiguous_format) → Tensor
+```
+
+Returns a tensor filled with uninitialized data on `device="tvarant"`.
+
+## Parameters
+
+- **size** (`int...`) – Size of the returned tensor.
+- **dtype** (`torch.dtype, optional`) – Desired data type. Defaults to the global default (prefer `float32`).
+- **device** (`torch.device, optional`) – Use `"tvarant"` / `"tvarant:0"`.
+- **memory_format** (`torch.memory_format, optional`) – Memory format for the result.
+
+## Shape
+
+- Output: $(\textit{size})$
+
+## Example
+
+```python
+x = torch.empty(2, 3, device='tvarant', dtype=torch.float32)
+```
+
+---
+
+[← All supported ops](index.md)

docs/api/ops/index.md

@@ -0,0 +1,42 @@
+# Supported ops
+
+ATen and custom ops registered for `PrivateUse1` (`device="tvarant"`).
+Unlisted ops **fall back to CPU** (tensors are copied as needed).
+
+Click an op for signature, parameters, shape, and examples.
+
+## Factories & memory
+
+| Op | Page |
+|---|---|
+| `empty` / `empty_strided` | [empty](empty.md) |
+| `zeros` / `ones` / `full` | [zeros / ones / full](zeros.md) |
+| `clone` / `contiguous` | [clone / contiguous](clone.md) |
+
+## Pointwise
+
+| Op | Page |
+|---|---|
+| `add` | [add](add.md) |
+| `mul` | [mul](mul.md) |
+| `sub` | [sub](sub.md) |
+| `relu` | [relu](relu.md) |
+| `silu` | [silu](silu.md) |
+| `neg` / `abs` / `exp` / `sqrt` / … | [unary](neg.md) |
+
+## Linear algebra & NN
+
+| Op | Page |
+|---|---|
+| `mm` / `addmm` / `matmul` / `bmm` | [mm family](mm.md) |
+| `linear` | [linear](linear.md) |
+| `softmax` | [softmax](softmax.md) |
+| `layer_norm` | [layer_norm](layer_norm.md) |
+| `embedding` | [embedding](embedding.md) |
+
+## Custom (`torch.ops.tvarant`)
+
+| Op | Page |
+|---|---|
+| `linear_act` | [linear_act](linear_act.md) |
+| `pointwise` | [pointwise](pointwise.md) |