innards 0.0.1__tar.gz

innards-0.0.1/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

innards-0.0.1/PKG-INFO

@@ -0,0 +1,401 @@
+Metadata-Version: 2.4
+Name: innards
+Version: 0.0.1
+Summary: See inside open-weight models while they run on your machine: read-only hooks on the model you loaded; for MoE models, which experts each token and topic activates during inference and training, and the expert set a topic needs; one analyzer per architectural feature; KV and memory predicted vs measured; a live web view.
+Keywords: llm,kv-cache,transformers,attention,gqa,mla,deltanet,mixture-of-experts,moe,expert-routing,interpretability,observability,profiling,inference
+Author: Jai
+License-Expression: Apache-2.0
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Programming Language :: Python :: 3.15
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: System :: Monitoring
+Classifier: Typing :: Typed
+Requires-Dist: httpx>=0.28.1
+Requires-Dist: huggingface-hub>=1.5
+Requires-Dist: innards[hf,view,aws,docs] ; extra == 'all'
+Requires-Dist: boto3>=1.34 ; extra == 'aws'
+Requires-Dist: pypdf>=5.0 ; extra == 'docs'
+Requires-Dist: psutil>=5.9 ; extra == 'hf'
+Requires-Dist: torch>=2.4 ; extra == 'hf'
+Requires-Dist: transformers>=5.0 ; extra == 'hf'
+Requires-Dist: fastapi>=0.115 ; extra == 'view'
+Requires-Dist: uvicorn>=0.30 ; extra == 'view'
+Requires-Python: >=3.10
+Project-URL: Homepage, https://pypi.org/project/innards/
+Project-URL: Documentation, https://pypi.org/project/innards/#description
+Provides-Extra: all
+Provides-Extra: aws
+Provides-Extra: docs
+Provides-Extra: hf
+Provides-Extra: view
+Description-Content-Type: text/markdown
+
+<h1 align="center">Innards</h1>
+<h3 align="center">See inside an open-weight model while it runs on your machine: which MoE experts every token and topic activates, and what each architectural feature does</h3>
+
+<p align="center">
+  <a href="https://medium.com/@jaiganesan"><img alt="Medium: @jaiganesan" src="https://img.shields.io/badge/Medium-@jaiganesan-000000?style=for-the-badge&logo=medium&logoColor=white"></a>
+  <a href="https://www.linkedin.com/in/jaiganesan-n/"><img alt="LinkedIn: jaiganesan-n" src="https://img.shields.io/badge/LinkedIn-jaiganesan--n-0A66C2?style=for-the-badge&logo=linkedin&logoColor=white"></a>
+</p>
+<p align="center">
+  <img alt="Version 0.0.1" src="https://img.shields.io/badge/version-0.0.1-555555?style=flat-square">
+  <img alt="Code license: Apache 2.0" src="https://img.shields.io/badge/code-Apache%202.0-blue?style=flat-square">
+  <img alt="Python 3.10 to 3.15" src="https://img.shields.io/badge/Python-3.10%20to%203.15-3776AB?style=flat-square&logo=python&logoColor=white">
+  <img alt="PyTorch 2.4 or later" src="https://img.shields.io/badge/PyTorch-2.4%2B-EE4C2C?style=flat-square&logo=pytorch&logoColor=white">
+  <img alt="Transformers 5" src="https://img.shields.io/badge/Transformers-5-FFD21E?style=flat-square&logo=huggingface&logoColor=black">
+  <img alt="Runs on a MacBook Air M5 (MPS), 16 GB" src="https://img.shields.io/badge/runs%20on-MacBook%20Air%20M5%20(MPS)%2C%2016%20GB-555555?style=flat-square&logo=apple&logoColor=white">
+</p>
+
+<p align="center">
+  <img alt="Granite 3.1 1B-A400M runs unchanged; Innards reads which experts each token chose in every MoE layer and counts them per topic" src="docs/figures/banner_light.png" width="100%">
+</p>
+
+Measuring and analysis code for one question: **which experts of a Mixture-of-Experts model does a token, a topic or a task actually use, measured while the model runs and without changing what it computes?** Your code loads the original checkpoint and runs it, for inference or for training; Innards attaches read-only hooks to that model and analyzes what they see beside it. For every model, MoE or not, it also records tokens, context, KV cache, memory and timing, and serves a live web view from the same machine. It targets models that run on a 16 GB machine at 16-bit precision or higher, through Hugging Face Transformers 5.
+
+> [!NOTE]
+> **Objective.** I am researching whether the experts that a specialized topic, task or token activates can be taken out of an MoE model and run on their own, as a smaller MoE with the model's own structure, without the full model. That starts with a measurement: which experts every token chose, in every layer, per topic, during inference and training. Innards does that step and exports the expert set a topic needs; building the smaller model is the job of your own code. Innards never removes, merges or edits experts.
+
+| Contents | |
+| --- | --- |
+| [1. What Innards does](#1-what-innards-does) | The parts, install |
+| [2. MoE expert routing](#2-moe-expert-routing) | Trace topics in inference and training, what is measured and derived, any MoE architecture |
+| [3. Observe a model](#3-observe-a-model) | One turn per `generate()`: KV, memory, timing, feature analyzers |
+| [4. The web view](#4-the-web-view) | Conversation, Analysis and Experts tabs |
+| [5. Where results go](#5-where-results-go) | Run folders, schema 1.0 |
+| [6. Checks, lineup and KV](#6-checks-lineup-and-kv) | Before loading, pairs of models, KV from a config |
+| [Run it](#run-it) | From a clone: setup and tests; repository layout |
+
+---
+
+## 1. What Innards does
+
+| Part | What it gives you |
+| --- | --- |
+| MoE expert routing | Which experts every token and topic activates, in inference and training, on any MoE architecture (including remote code such as `facebook/MobileMoE`), and the experts a topic needs per layer, ready to export |
+| Every model | Tokens, context, KV cache per layer (predicted from the config and measured), memory and timing |
+| Feature analyzers | One per architectural feature (Gated DeltaNet, gated attention, MoE, mHC, CSA/HCA, DSA, MLA), picked from the model's config |
+| Web view | Served from the machine that runs the model: no central site, no API key |
+
+Innards never loads or runs a model itself, and its hooks only read. Missing measurements are `null`, never zero, and every value computed from a measurement is labeled as derived.
+
+### Install
+
+<!-- test: install (checked by the fresh-environment wheel install in CI) -->
+```bash
+pip install innards                  # core: analysis of stored traces, KV calculator, lineage
+pip install "innards[hf]"            # + torch, transformers: observe and trace a model
+pip install "innards[view]"          # + the web view
+pip install "innards[all]"           # everything, including S3 sync and PDF documents
+
+uv add "innards[hf,view]"            # inside a uv project
+```
+
+---
+
+## 2. MoE expert routing
+
+### 2.1 Trace topics during inference
+
+`innards.experts.trace` hooks every MoE layer of the model you loaded. Inside `trace.topic(...)`, every forward pass is recorded (the prefill and each decoded token); outside a topic nothing is.
+
+<!-- test: requires=hf,network -->
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+import innards
+
+model_id = "ibm-granite/granite-3.1-1b-a400m-instruct"  # 32 experts, 8 per token; 2.7 GB in bf16
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16)
+
+trace = innards.experts.trace(model, tokenizer, run="topic-experts")  # read-only hooks on every MoE layer
+prompts = {
+    "math": ["The derivative of x squared is", "Solve for x: 3x + 5 = 20. First,"],
+    "code": ["def fibonacci(n):\n    if n <", "import numpy as np\narr = np.zeros("],
+}
+for topic, texts in prompts.items():
+    for text in texts:
+        with trace.topic(topic):
+            model.generate(**tokenizer(text, return_tensors="pt"), max_new_tokens=16, do_sample=False)
+
+experts = trace.map()  # topic x layer x expert, from the routing the model chose
+print(experts.top("math", layer=12, n=4))  # the experts math tokens chose most in layer 12
+print(experts.specialists("code", n=3))  # chosen far more for code than for math
+keep = experts.select("math", coverage=0.9)  # per layer, the fewest experts that take 90% of math's routing
+keep.save("math-experts.json")
+print(keep)
+
+routing = trace.tokens()  # every recorded token's experts, per layer
+print(routing.token(" x")["counts"][12])  # how often each expert took the token " x" in layer 12
+print(routing.expert(layer=12, expert=keep.keep(12)[0])[:3])  # the tokens one expert takes most
+trace.close()  # removes the hooks and writes innards_runs/topic-experts/experts/<model>/
+```
+
+### 2.2 During training
+
+Give one topic per sequence of the batch. Padding is never counted, and the forward that gradient checkpointing recomputes inside `backward()` is not counted twice. The example fine-tunes only the routers, to stay light; the trace works the same when every weight trains.
+
+<!-- test: requires=hf,network -->
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+import innards
+
+model_id = "ibm-granite/granite-3.1-1b-a400m-instruct"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).train()
+model.requires_grad_(False)
+routers = [m for name, m in model.named_modules() if name.endswith(".router")]
+for router in routers:
+    router.requires_grad_(True)
+optimizer = torch.optim.SGD([p for r in routers for p in r.parameters()], lr=1e-5)
+
+trace = innards.experts.trace(model, tokenizer, run="finetune", tokens=False)  # counts only: light for long runs
+batch = tokenizer(
+    ["The integral of 2x is x squared.", "for i in range(3):\n    print(i)"], return_tensors="pt", padding=True
+)
+labels = batch["input_ids"].masked_fill(batch["attention_mask"] == 0, -100)
+with trace.topic(["math", "code"]):  # one label per sequence
+    loss = model(**batch, labels=labels).loss
+loss.backward()
+optimizer.step()
+print(trace.map())  # <ExpertMap ...: 24 MoE layers; tokens math: ..., code: ...>
+trace.close()
+```
+
+### 2.3 What is measured, and what is derived
+
+| | Value | Meaning |
+| --- | --- | --- |
+| Measured | `tokens` | Tokens of a topic that went through each MoE layer (padding excluded) |
+| Measured | `counts` | How often each expert was in a token's top-k |
+| Measured | per-token routing | Each token's chosen experts and their weights, per layer (`tokens=True`) |
+| Derived | `selection` | Share of the topic's tokens whose top-k included the expert (`counts / tokens`) |
+| Derived | `gate` | The expert's average share of a token's routing weight |
+| Derived | `lift` | Selection for this topic ÷ selection for every other topic together |
+| Derived | `select(topic, coverage)` | Per layer, the fewest experts that together take `coverage` of the topic's routing |
+
+`select` saves the expert ids to keep per layer, with the module paths of each layer's router, experts and shared expert and the routing settings from the config. That file is the input for your own code that builds the smaller model; shared experts run for every token and always stay.
+
+### 2.4 Any MoE architecture
+
+Each MoE layer is read from the first place that carries the model's own routing decision:
+
+| Source | Used when |
+| --- | --- |
+| `experts_input` | The block calls `experts(hidden, top_k_index, top_k_weights)`: transformers 5 MoE models (Qwen3-MoE, Qwen3.5-MoE, Mixtral, OLMoE, Granite MoE, GPT-OSS, DeepSeek-V2/V3, Phi-MoE, LFM2-MoE, Nemotron-H ...) |
+| `router_output` | The router module returns the chosen experts |
+| `topk` | Code that routes inline, such as `facebook/MobileMoE`: the block's own `torch.topk` |
+
+MoE checkpoints that fit 16 GB at bf16: `ibm-granite/granite-3.1-1b-a400m-instruct` (2.7 GB), `facebook/MobileMoE-S-Base` (2.4 GB, remote code), `ibm-granite/granite-swash-3b-a600m` (6 GB), `microsoft/Phi-tiny-MoE-instruct` (7.5 GB), `arcee-ai/Trinity-Nano-Preview` (12.2 GB) and `allenai/OLMoE-1B-7B-0125-Instruct` (13.8 GB, tight).
+
+From the command line, `innards experts` lists every trace under `./innards_runs`, and the **Experts** tab of the web view shows a trace live:
+
+<!-- test: cli -->
+```bash
+innards experts                                   # every trace under ./innards_runs: model, topics, tokens
+```
+
+<!-- test: text -->
+```text
+innards experts innards_runs/topic-experts --topic math                 # top experts per layer
+innards experts innards_runs/topic-experts --topic code --by lift       # specialists against the other topics
+innards experts innards_runs/topic-experts --topic math --select 0.9 --out math-experts.json
+```
+
+---
+
+## 3. Observe a model
+
+`innards.observe` records one turn per `generate()`: tokens, context, KV per layer (predicted vs measured), memory and timing, plus an analyzer for each feature the config says the model has.
+
+<!-- test: requires=hf,network -->
+```python
+import innards
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-0.8B")
+model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-0.8B")  # your code loads the original checkpoint
+
+watch = innards.observe(model, tokenizer, model_id="Qwen/Qwen3.5-0.8B", run="deltanet-study")
+
+messages = [{"role": "user", "content": "Remember this: my name is Priya."}]
+inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)
+out = model.generate(**inputs, max_new_tokens=32)  # your code, unchanged: one turn
+
+turn = watch.last_turn().record
+print(turn.kv.measured_bytes == turn.kv.predicted_bytes)  # True
+print(sorted(turn.features))  # ['gated_attention', 'gated_deltanet', 'kv']
+watch.close()
+```
+
+| Analyzer | Shows |
+| --- | --- |
+| `kv` (every model) | KV per layer and layer type, growth, context, predicted vs measured |
+| `gated_deltanet` | Write strength, retention, half-life, fixed state next to growing KV |
+| `gated_attention` | Gate openness per head and token |
+| `moe` | Expert load per turn, routing entropy |
+| `mhc` | Stream mixing, doubly-stochastic check, signal gain across depth |
+| `csa_hca`, `dsa`, `mla` | Compressed and indexer caches, latent vs expanded layout |
+
+`observe` and `innards.experts.trace` can watch the same model at once.
+
+---
+
+## 4. The web view
+
+`innards view` (or `watch.view()`) serves a monitoring page from the machine that runs the model:
+
+| Tab | What it shows |
+| --- | --- |
+| Conversation | Run a conversation set turn by turn, or type your own messages; tokens reused, KV change and speed per reply |
+| Analysis | Each feature at work, KV/context/memory per turn predicted vs measured, and the same set compared with the version before |
+| Experts | Experts per layer for a topic, topics side by side, the experts each topic needs, specialists, each sequence as tokens × experts; downloads the expert set |
+
+<!-- test: text -->
+```text
+innards view                                   # ./innards_runs on http://127.0.0.1:8000
+innards view ./innards_runs/deltanet-study     # one run folder, e.g. copied from another machine
+ssh -L 8000:127.0.0.1:8000 user@<gpu-host>     # a remote machine's view through a tunnel
+```
+
+---
+
+## 5. Where results go
+
+Results stay on the machine that runs the model:
+
+<!-- test: text -->
+```text
+innards_runs/<run>/run.json                                 name, host, environment
+innards_runs/<run>/<conversation>/turns.jsonl               one TurnRecord (schema 1.0) per turn
+innards_runs/<run>/experts/<name>/summary.json              expert trace: topics, tokens and counts per layer
+innards_runs/<run>/experts/<name>/tokens/shard-*.safetensors   per-token experts, weights, positions, topics
+```
+
+<!-- test: check-output -->
+```python
+from innards.schema import TurnRecord
+
+record = TurnRecord.from_dict(
+    {
+        "schema_version": "1.0",
+        "run": "deltanet-study",
+        "conversation": "20261103-081500-Qwen3.5-0.8B-a1b2",
+        "turn": 7,
+        "timestamp": "2026-11-03T08:15:42Z",
+        "model": {"id": "Qwen/Qwen3.5-0.8B", "architecture": "hybrid", "features": ["gated_deltanet"]},
+        "runtime": {"backend": "observe"},
+        "kv": {"predicted_bytes": 2154496, "measured_bytes": 2154496, "measured_source": "hf_cache"},
+    }
+)
+print(record.kv.measured_bytes, record.memory.peak_bytes)  # 2154496 None
+```
+
+---
+
+## 6. Checks, lineup and KV
+
+Before anything is downloaded, `validate` checks architecture support, the device against the checkpoint's published precision, and whether weights plus KV fit the memory budget (16 GB by default, `INNARDS_GPU_BUDGET_GB`):
+
+<!-- test: check-output requires=hf -->
+```python
+from innards.validate import validate
+
+report = validate("ibm-granite/granite-swash-3b-a600m", device="mps", check_access=False)
+print(report.status)  # Ready
+small = validate("ibm-granite/granite-swash-3b-a600m", device="mps", budget_gb=4, check_access=False)
+print([check.id for check in small.blocked])  # ['memory']
+```
+
+The lineup pairs a newer model with the version before it, each running on a 16 GB machine at 16-bit precision:
+
+| Feature | Newer model | Version before |
+| --- | --- | --- |
+| Gated DeltaNet + gated attention | Qwen3.5-0.8B (or 2B) | Qwen3-0.6B (or 1.7B) |
+| MoE with a shared expert + sliding-window attention | Granite Swash 3B-A600M | Granite 3.1 3B-A800M |
+
+`predict` gives KV bytes and a per-layer breakdown from a config alone, and `explain` says what an architecture changed from its predecessor:
+
+<!-- test: check-output -->
+```python
+from innards.kv import predict
+from innards.lineage import explain
+from innards.lineage.lineup import model_config
+
+p = predict(model_config("ibm-granite/granite-swash-3b-a600m"), context_tokens=32768, kv_dtype="bf16")
+print(p.architecture, round(p.total_bytes / 2**20))  # sliding_window 258
+print(explain("gqa").predecessor, "->", explain("gqa").family)  # mha -> gqa
+```
+
+<!-- test: cli -->
+```bash
+innards lineup                                               # the pairs and what each newer model adds
+innards predict --model Qwen/Qwen3-0.6B --context 32768 --kv-dtype fp8
+innards explain sliding_window
+innards validate Qwen/Qwen3.5-0.8B --device cpu --set with_context
+```
+
+---
+
+## Run it
+
+From a clone, with [uv](https://docs.astral.sh/uv/):
+
+<!-- test: dev -->
+```bash
+git clone https://github.com/jaiganesan-n/innards.git && cd innards
+uv sync --all-extras                                    # Python 3.13, torch, transformers 5, FastAPI, dev tools
+uv run pytest                                           # unit tests, README examples, small real checkpoints
+uv run ruff check && uv run ruff format --check && uv run mypy src
+INNARDS_REMOTE_CODE_TESTS=1 uv run pytest tests/test_experts_models.py -k mobilemoe   # facebook/MobileMoE (remote code)
+uv run --with playwright python docs/figures/src/make_banner.py --measure           # re-trace Granite, redraw the banner
+```
+
+### Repository layout
+
+<!-- test: text -->
+```text
+innards/
+├── README.md, LICENSE
+├── pyproject.toml, uv.lock, .python-version   # uv; Python 3.10 to 3.15
+├── src/innards/
+│   ├── experts/     routing.py, trace.py, analysis.py     # MoE expert routing
+│   ├── observer/    watch.py, probes.py, observer.py, cache.py
+│   ├── features/    one analyzer per architectural feature
+│   ├── kv/          KV calculator
+│   ├── lineage/     lineup, pinned configs, architecture families
+│   ├── view/        web view (FastAPI, static page)
+│   ├── schema.py, store.py, sync.py, validate.py, sets.py
+│   └── cli.py       # innards <command>
+├── tests/           replicas.py (tiny random-weight models), test_*.py
+├── scripts/         overhead_bench.py
+├── docs/figures/    banner_light.png, banner_data.json (what the banner shows), src/make_banner.py
+└── .github/workflows/  ci.yml, publish.yml (PyPI on a v* tag)
+```
+
+## License
+
+Code: Apache 2.0 (`LICENSE`). Models keep their own terms: Innards never downloads model weights itself (`facebook/MobileMoE` is gated, for non-commercial research only).
+
+---
+
+## Author
+
+**Jai Ganesan** · [Medium](https://medium.com/@jaiganesan) · [LinkedIn](https://www.linkedin.com/in/jaiganesan-n/)
+
+> [!IMPORTANT]
+> **Note.** Innards is developed on one MacBook Air M5 with 16 GB of memory, so it targets models that fit there. If you trace another MoE model, or one whose routing Innards does not read yet, let me know.