Update documentation links and remove an old requirements file.

.github/workflows/utils/setup_runner.sh

@@ -14,8 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Heavily influenced by
-# https://github.com/openxla/iree/tree/main/build_tools/github_actions/runner/config
+# Not currently used after the move to ML Actions (GKE). Kept around in case
+# we need to reference it for older workflows or as a guide for future workflow
+# configurations.
 
 # This file sets up a tpu vm to be used as a github runner for testing. 
 # It creates a user runner without sudo permissions to

README.md

@@ -41,7 +41,7 @@ See our installation guide to [install MaxText with pip](https://maxtext.readthe
 * \[September 15, 2025\] MaxText is now available as a [PyPI package](https://pypi.org/project/maxtext). Users can now [install maxtext through pip](https://maxtext.readthedocs.io/en/latest/guides/install_maxtext.html).
 * \[September 5, 2025\] MaxText has moved to an `src` layout as part of [RESTRUCTURE.md](RESTRUCTURE.md). For existing environments, please run `pip install -e .` from MaxText root.
 * \[August 13, 2025\] The Qwen3 2507 MoE family of models is now supported: MoEs: 235B Thinking & 280B Coder as well as existing dense models: 0.6B, 4B, 8B, 14B, and 32B.  
-* \[July 27, 2025\] Updated TFLOPS/s calculation ([PR](https://github.com/AI-Hypercomputer/maxtext/pull/1988)) to account for causal attention, dividing the attention flops in half. Accounted for sliding window and chunked attention reduced attention flops in [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2009) and [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2030). Changes impact large sequence configs, as explained in this [doc](https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/guides/performance_metrics.md)  
+* \[July 27, 2025\] Updated TFLOPS/s calculation ([PR](https://github.com/AI-Hypercomputer/maxtext/pull/1988)) to account for causal attention, dividing the attention flops in half. Accounted for sliding window and chunked attention reduced attention flops in [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2009) and [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2030). Changes impact large sequence configs, as explained in this [doc](https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/explanations/performance_metrics.md)  
 * \[July 16, 2025\] We will be restructuring the MaxText repository for improved organization and clarity. Please review the [proposed structure](https://github.com/AI-Hypercomputer/maxtext/blob/main/RESTRUCTURE.md) and provide feedback.  
 * \[July 11, 2025\] Multi-Token Prediction (MTP) training support\! Adds an auxiliary loss based on predicting multiple future tokens, inspired by [DeepSeek-V3 paper](https://arxiv.org/html/2412.19437v1), to enhance training efficiency.  
 * \[June 25, 2025\] DeepSeek R1-0528 variant is now supported.  

docs/explanations/jax_ai_libraries_chosen.md

@@ -60,7 +60,7 @@ For more information on using Orbax, please refer to https://github.com/google/o
 
 Its APIs are explicitly designed for the multi-host paradigm, simplifying the process of ensuring that each host loads a unique shard of the global batch.
 
-For more information on using Grain, please refer to https://github.com/google/grain and the grain guide in maxtext located at https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/guides/data_input_grain.md 
+For more information on using Grain, please refer to https://github.com/google/grain and the grain guide in maxtext located at https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/guides/data_input_pipeline/data_input_grain.md 
 
 ## Qwix: For native JAX quantization
 

docs/guides/checkpointing_solutions/multi_tier_checkpointing.md

@@ -43,7 +43,7 @@ Multi-tier checkpointing stores checkpoints across multiple tiers of storage:
 
 To run workloads with Multi-Tier Checkpointing (MTC), you need a Google Kubernetes Engine (GKE) cluster with the necessary drivers and features enabled. You can create a properly configured cluster using the **XPK** or by setting it up manually with `gcloud` commands following [Google Cloud Checkpointing Documentation](https://cloud.google.com/kubernetes-engine/docs/how-to/machine-learning/training/multi-tier-checkpointing).
 
-The [xpk script](https://github.com/AI-Hypercomputer/xpk/blob/develop/xpk.py) provides a streamlined way to create a GKE cluster with all the required MTC settings. The key flags used are:
+The [xpk script](https://github.com/AI-Hypercomputer/xpk/blob/main/xpk.py) provides a streamlined way to create a GKE cluster with all the required MTC settings. The key flags used are:
 
 | Flag | Description |
 | :--- | :--- |

docs/guides/understand_logs_and_metrics.md

@@ -174,7 +174,7 @@ This log shows that each of the four TPUs has `95.74 GB` of available High Bandw
 ### 3.3. Model TFLOP per device
 
 The **model FLOPs** are the floating point operations to perform model computation. For training, the computation includes a single forward and backward pass.
-- In MaxText, we estimate model FLOPs by summing operations in matrix multiplications (matmuls); see [calculate_tflops_training_per_device](https://github.com/AI-Hypercomputer/maxtext/blob/28e5097ac467ed8b1d17676d68aa5acc50f9d60d/src/MaxText/maxtext_utils.py#L480).
+- In MaxText, we estimate model FLOPs by summing operations in matrix multiplications (matmuls); see [calculate_tflops_training_per_device](../../src/MaxText/maxtext_utils.py#L480).
 - The number of model FLOPs is dependent on model architecture, input size (batch size, sequence length), and gradient accumulation steps. It does not include optimization operations.
 - We break down the FLOPs into two parts:
   - "Learnable weight FLOPs" are matmuls between activations and learnable weights. Specifically, this occurs in embedding, feed forward networks, attention-related projections, and unembedding.

docs/index.md

@@ -30,7 +30,7 @@ Check out our [Read The Docs site](https://maxtext.readthedocs.io/en/latest/) or
 
 * [September 5, 2025] MaxText has moved to an `src` layout as part of [RESTRUCTURE.md](https://github.com/AI-Hypercomputer/maxtext/blob/main/RESTRUCTURE.md).
 * [August 13, 2025] The Qwen3 2507 MoE family of models is now supported: MoEs: 235B Thinking & 280B Coder as well as existing dense models: 0.6B, 4B, 8B, 14B, and 32B.
-* [July 27, 2025] Updated TFLOPS/s calculation ([PR](https://github.com/AI-Hypercomputer/maxtext/pull/1988)) to account for causal attention, dividing the attention flops in half. Accounted for sliding window and chunked attention reduced attention flops in [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2009) and [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2030). Changes impact large sequence configs, as explained in this [doc](https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/guides/performance_metrics.md)
+* [July 27, 2025] Updated TFLOPS/s calculation ([PR](https://github.com/AI-Hypercomputer/maxtext/pull/1988)) to account for causal attention, dividing the attention flops in half. Accounted for sliding window and chunked attention reduced attention flops in [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2009) and [PR](https://github.com/AI-Hypercomputer/maxtext/pull/2030). Changes impact large sequence configs, as explained in this [doc](explanations/performance_metrics.md)
 * [July 16, 2025] We will be restructuring the MaxText repository for improved organization and clarity. Please review the [proposed structure](https://github.com/AI-Hypercomputer/maxtext/blob/main/RESTRUCTURE.md) and provide feedback.
 * [July 11, 2025] Multi-Token Prediction (MTP) training support\! Adds an auxiliary loss based on predicting multiple future tokens, inspired by [DeepSeek-V3 paper](https://arxiv.org/html/2412.19437v1), to enhance training efficiency.
 * [June 25, 2025] DeepSeek R1-0528 variant is now supported

docs/reference/tiering.md

@@ -40,4 +40,4 @@ For each of the TPU platforms listed below, we present a list of optimized model
 | Mixtral 8X7B | [Link](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/v5p/Mixtral-8X7B-Maxtext) | 256 Chips(8x4x4), bf16, SL=4096 | 52.56% | 2,909 |
 
 [^1]:  Performance results are subject to variations based on system configuration, software versions, and other factors. These benchmarks represent point-in-time measurements under specific conditions.
-[^2]:  Some older TFLOPS/s results are impacted by an updated calculation for causal attention ([PR #1988](https://github.com/AI-Hypercomputer/maxtext/pull/1988)), which halves the attention FLOPs. This change particularly affects configurations with large sequence lengths. For more details, please refer to the [performance metrics guide](https://github.com/AI-Hypercomputer/maxtext/blob/main/docs/guides/performance_metrics.md).
+[^2]:  Some older TFLOPS/s results are impacted by an updated calculation for causal attention ([PR #1988](https://github.com/AI-Hypercomputer/maxtext/pull/1988)), which halves the attention FLOPs. This change particularly affects configurations with large sequence lengths. For more details, please refer to the [performance metrics guide](performance-metrics).

docs/tutorials/full_finetuning.md

@@ -76,7 +76,7 @@ These scripts can provide a reference point for various scripts.
 
 ### MaxText checkpoint to Hugging Face
 
-Post finetuning or pre-training, MaxText also provides scripts to convert MaxText format weights back to [Hugging Face](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/MaxText/llama_mistral_mixtral_orbax_to_hf.py).
+Post finetuning or pre-training, MaxText also provides scripts to convert MaxText format weights back to [Hugging Face](../../src/MaxText/utils/ckpt_scripts/llama_mistral_mixtral_orbax_to_hf.py).
 
 #### Dataset