fix(docs): correct image path (#4268)

* fix(docs): correct image path for e2e_flow_part1.png in pretraining tutorial

* fix image path in finetuning.rst

* fix image path in serving.rst

* fix image path in sparsity.rst

* Fix image path for loss curves in pretraining.rst

Author: itsPronay

docs/source/contributing/sparsity.rst

@@ -53,7 +53,7 @@ If you can get your dense matrix into a **2:4 sparse format**, we can speed up m
 This also allows users with existing sparse weights in a dense format to take advantage of our fast sparse kernels. We anticipate many users to come up with their own custom frontend masking solution or to use another third party solution, as this is an active area of research.
 
 
-.. image:: ../static/pruning_ecosystem_diagram.png
+.. image:: ../../static/pruning_ecosystem_diagram.png
    :alt: pruning_flow
 
 
@@ -109,7 +109,7 @@ In order to avoid confusion, we generally try to use sparsity to refer to tensor
 Roughly, the flow for achieving a more performant pruned model looks like this:
 
 
-.. image:: ../static/pruning_flow.png
+.. image:: ../../static/pruning_flow.png
    :alt: flow
 
 

docs/source/eager_tutorials/finetuning.rst

@@ -7,7 +7,7 @@ techniques integrated into our partner frameworks. This is part 2 of 3
 such tutorials showcasing this end-to-end flow, focusing on the
 fine-tuning step.
 
-.. image:: ../static/e2e_flow_part2.png
+.. image:: ../../static/e2e_flow_part2.png
 
 Fine-tuning is an important step for adapting your pre-trained model
 to more domain-specific data. In this tutorial, we demonstrate 3 model

docs/source/eager_tutorials/pretraining.rst

@@ -7,7 +7,7 @@ techniques integrated into our partner frameworks. This is part 1 of 3
 such tutorials showcasing this end-to-end flow, focusing on the
 pre-training step.
 
-.. image:: ../static/e2e_flow_part1.png
+.. image:: ../../static/e2e_flow_part1.png
 
 Pre-training with float8 using torchao can provide `up to 1.5x speedups <https://pytorch.org/blog/training-using-float8-fsdp2/>`__ on 512 GPU clusters,
 and up to `1.34-1.43x speedups <https://pytorch.org/blog/accelerating-large-scale-training-and-convergence-with-pytorch-float8-rowwise-on-crusoe-2k-h200s/>`__ on 2K H200 clusters with the latest `torchao.float8` rowwise recipe.
@@ -116,7 +116,7 @@ This is because rowwise scaling using a more granular scaling factor (per row, i
 
 Below you can see the loss curves comparing bfloat16, float8 tensorwise, and float8 rowwise training for training Llama3-8B on 8xH100 GPUs:
 
-.. image:: ../static/fp8-loss-curves.png
+.. image:: ../../static/fp8-loss-curves.png
   :alt: Loss curves for training Llama3-8B on 8xH100s with torchtitan using bfloat16, float8 tensorwise, and float8 rowwise training.
 
 

docs/source/eager_tutorials/serving.rst

@@ -3,7 +3,7 @@
 
 TorchAO provides an end-to-end pre-training, fine-tuning, and serving model optimization flow by leveraging our quantization and sparsity techniques integrated into our partner frameworks. This is part 3 of 3 such tutorials showcasing this end-to-end flow, focusing on the serving step.
 
-.. image:: ../static/e2e_flow_part3.png
+.. image:: ../../static/e2e_flow_part3.png
 
 This tutorial demonstrates how to perform post-training quantization and deploy models for inference using torchao as the underlying optimization engine, seamlessly integrated through HuggingFace Transformers, vLLM, and ExecuTorch.