Single source for images in content folders

Author: laithsakka

content/dynamic_shapes/2025-07-08-guard-free-dynamic-shapes.md

@@ -16,7 +16,7 @@ tags: [dynamic_shapes, unbacked, dde, guard_free, export, torch.compile]
 > achieving a **55% reduction** in framework DDEs and saving weeks of
 > engineering time on complex model exports.
 
-![DDE reduction progress](/devlogs/images/2025-07-08-dde-progress.jpg)
+![DDE reduction progress](./images/2025-07-08-dde-progress.jpg)
 
 ## Why are guard-free dynamic shape graphs necessary?
 

content/dynamic_shapes/2026-01-20-backed-to-unbacked.md

@@ -5,7 +5,7 @@ date: 2026-01-20
 tags: [dynamic_shapes, unbacked, backed, torch.compile, frontier]
 ---
 
-![Hint stamp](/devlogs/images/2026-01-20-backed-unbacked-header.jpg)
+![Hint stamp](./images/2026-01-20-backed-unbacked-header.jpg)
 
 > **TL;DR** – We expect unbacked dynamic shapes to become the dominant
 > shape mechanism for Frontier-style workloads due to their better

content/dynamic_shapes/2026-02-27-compile-time-unbacked-export.md

@@ -5,7 +5,7 @@ date: 2026-02-27
 tags: [dynamic_shapes, unbacked, torch.export, compile_time, symbolic_shapes]
 ---
 
-![Compile time reduction](/devlogs/images/2026-02-27-compile-time-header.jpg)
+![Compile time reduction](./images/2026-02-27-compile-time-header.jpg)
 
 > **TL;DR** – A regression report revealed that
 > exporting a model with many unbacked (data-dependent) symbols took
@@ -20,7 +20,7 @@ A report indicated a severe slowdown when exporting a model that heavily
 uses data-dependent operations (i.e., unbacked symbolic shapes).  Profiling
 showed that most of the time was spent inside the symbolic shape system.
 
-![Profiling hotspot: SymNode.expr takes significant trace time](/devlogs/images/2026-02-27-compile-time-fig1.png)
+![Profiling hotspot: SymNode.expr takes significant trace time](./images/2026-02-27-compile-time-fig1.png)
 
 At the time of investigation, `torch.export` did not support profiling
 out of the box, which made root-cause analysis difficult.  After
@@ -77,7 +77,7 @@ became a significant cost.
 A per-SymNode cache was added that is invalidated only when replacements
 change (tracked via a version counter).  **9% improvement.**
 
-![Profiling hotspot: SymNode.expr takes significant trace time](/devlogs/images/2026-02-27-compile-time-fig1.png)
+![Profiling hotspot: SymNode.expr takes significant trace time](./images/2026-02-27-compile-time-fig1.png)
 
 
 ### 4. Optimize `_smart_symbol_sort` hint access ([PR #174655](https://github.com/pytorch/pytorch/pull/174655))
@@ -96,7 +96,7 @@ side is an unbacked symbol and the other is a constant triggered expensive
 sympy evaluation.  Since unbacked symbols have no assumptions, there's
 nothing to simplify.  Using `evaluate=False` skips this unnecessary work.
 **15% improvement.**
-![cost of sympy_eq](/devlogs/images/2026-02-27-compile-time-fig2.png)
+![cost of sympy_eq](./images/2026-02-27-compile-time-fig2.png)
 
 ### 6. Avoid redundant `compute_hint()` calls during expression construction ([PR #174664](https://github.com/pytorch/pytorch/pull/174664))
 
@@ -105,7 +105,7 @@ When `SymNode` operations produce results with unavailable hints, passing
 for no reason — it will fail.  A sentinel value now explicitly indicates
 "hint unavailable, don't recompute."  **7% improvement.**
 
-![compute hint overhead](/devlogs/images/2026-02-27-compile-time-fig3.png)
+![compute hint overhead](./images/2026-02-27-compile-time-fig3.png)
 
 ### 7. Skip static eval for unbounded unbacked symbols ([PR #174652](https://github.com/pytorch/pytorch/pull/174652))
 
@@ -121,7 +121,7 @@ range `[0, ∞]` will be fast due to the earlier optimization.
 performing redundant computations.  Using `_from_args` directly with
 `is_commutative=True` saves **5% improvement.**
 
-![Cumulative improvement: 264s → 87s](/devlogs/images/2026-02-27-compile-time-fig4.png)
+![Cumulative improvement: 264s → 87s](./images/2026-02-27-compile-time-fig4.png)
 
 ## Process takeaways
 

content/dynamic_shapes/2026-03-25-unbacked-perf-parity.md

@@ -5,7 +5,7 @@ date: 2026-03-25
 tags: [dynamic_shapes, unbacked, performance, vllm, torchbench, inductor]
 ---
 
-![Backed vs unbacked perf parity](/devlogs/images/2026-03-25-perf-parity-header.jpg)
+![Backed vs unbacked perf parity](./images/2026-03-25-perf-parity-header.jpg)
 
 > **TL;DR** – Unbacked dynamic shapes had 2x–20% slowdowns on TorchBench
 > and ~30% regressions on vLLM.  We fixed the root causes — now unbacked

hugo.toml

@@ -6,7 +6,7 @@ title = 'PyTorch DevLog'
 [params]
   description = 'Developer technical notes for PyTorch'
 
-summaryLength = 30
+summaryLength = 60
 
 [markup]
   [markup.goldmark]

static/css/style.css

@@ -409,7 +409,7 @@ a.tag:hover {
   font-size: 0.85rem;
   color: var(--color-text-muted);
   margin-bottom: 0.5rem;
-  max-height: 12rem;
+  max-height: 18rem;
   overflow: hidden;
 }