(5) The July 2026 Release Waves — vLLM 0.25→0.26 and SGLang 0.5.15→0.5.16

1. The Whole Timeline

July 2026 was the most change-dense month for these two engines in years: both nearly synchronized a generation of architecture turnover, then nearly released their next versions on the same day.

Fig 1: The two product lines moved in almost perfect lockstep through July — evidence they face the same bottlenecks.

2. vLLM Evolution Detail

2.1 v0.25.0 (7/11–7/12): the architecture turnover

This was the biggest leap, four headline changes:

Change	PR	Meaning
Model Runner V2 becomes default for all dense models	#39337	async-first, zero CPU-GPU sync, overlap step N and N+1
PagedAttention removed	#47361	abstraction pushed down to the attention-backend kernel; the paging mechanism itself is kept
Transformers v4 deprecated	#40389	must migrate to v5
Transformers backend parity	—	any new architecture HF implements is served day-0 at full speed

Also in tow: unified streaming-parse engine (#46610), native DSpark speculative decoding for DeepSeek V4 Pro (8×B300 ≈ 250 tok/s, 12–42% above MTP), heterogeneous-vocab general speculation (#38174), thinking-budget-aware speculation (#34668), and the compile requirement raised to C++20.

Performance character: MRv2's gains show most in small/medium batch — exactly the regime real Agent traffic lives in. If you only run large-batch offline jobs, you'll feel it far less.

2.2 v0.25.1 (7/14): a two-commit must-have patch

Two days after release came a patch, because a silent correctness bug was found:

#48330 mixed-dtype quant fusion guard: FlashInfer’s allreduce + RMSNorm + static-quant fused kernel, when the activation is BF16 and the RMSNorm weight is FP32, hits a dtype mismatch that reads 4-bit NVFP4 with the wrong bit pattern → corrupts hidden states → output degrades into repeated !!!!!. The fix adds a dtype-match guard: mismatched takes the safe path, matched keeps the fusion.
#47888: TorchCodec no longer blocks startup when FFmpeg is missing.

For anyone serving NVFP4 (Gemma4 / Qwen-family / GLM-5.2, etc.), 0.25.0 must upgrade to 0.25.1. This bug doesn't error or crash — it just quietly outputs garbage, so it's easy to miss in production.

2.3 7/18–7/25 main: stability wrap-up

0.25 was a big change, and a wave of edge-config fixes followed:

PR	Fixes
#48524	DFlash layer sizing wrong when `num_target_layers ≠ num_hidden_layers`
#49302	DSA crash under interruptible segmented CUDA Graph
#48843	`graph_pool_id` not set before full CUDA Graph capture
#49306	FA4 JIT warmup MLA fallback
#48860	KV Connector delayed request double-counts prefix-cache metric
#49292	Qwen3-VL M-RoPE on the Transformers backend
#49190	Cosmos3 Edge video model
#48816	GPTQ Qwen3.5 MTP weight-loading error when speculation on
#42569	FA4 on SM100 (Blackwell) adds FP8 KV cache support
#48683	ROCm up to AITER v0.1.16.post5
#45991	XPU adds DeepSeek-V4 `fuse_index_q` SYCL path
#49244	remove old partial-prefill params
#48914 / #49427	FlashInfer up 0.6.15; restore `dequant_cache` OOB guard

vLLM main merged roughly 64 commits/day in June. That speed means: production must lock release tags, don't follow main; run a full regression before upgrading.

2.4 v0.26.0 (7/25)

grammar failure no longer crashes the engine: a structured-output parse failure becomes a single-request error, not affecting other requests;
prefix-cache hit-rate reporting: observability filled in;
NVFP4 requires FlashInfer: dependency made explicit;
some startup flags renamed: check your launch script on upgrade.

2.5 7/29–7/31: shifting to production deployment

As the version cadence slowed, the focus turned to landing:

GLM-5.2 production deployment plan: NVFP4 + MTP + P/D combo. IndexerCache lifts MTP acceptance; PCP (context parallel) lifts prefill throughput from 20.1k to 27.3k;
P/D two-tier cache TieringManager (PR #42285);
async KV-load prefetch fix (PR #46694);
week-27 weekly-report theme: PD disaggregation via ZMQ + NIXL, Rust protocol refactor.

3. SGLang Evolution Detail

3.1 v0.5.15 (7/10): zero-overhead Spec V2

Feature	Note
Zero-overhead Spec V2 default	draft-extend made CUDA-graph-capturable, cut D2H/H2D, fused metadata → end-to-end +~11% TPS
IndexShare MTP	draft step reuses target model’s DSA top-k index, long-context draft cost down ~1.9×
Breakable CUDA Graph default	graph interruptible, balancing graph gains with scheduling flexibility
MLA context parallel decoding	context-parallel decoding for DeepSeek-family MLA
FlashInfer all-to-all MoE routing	MoE routing via FlashInfer
Native web search	built-in retrieval tool

It also fixed the 0.5.5–0.5.12 multimodal path-traversal vulnerability (GHSA-qwrp-wghp-94q2).

3.2 v0.5.15.post1 (7/14): GLM-5.2 tuning

Coinciding with the lmsys official blog “GLM-5.2 NVFP4 500 TPS”:

Spec V2 default +11% TPS
IndexShare MTP reuses top-k → ~1.9× cost cut
TopK-V2 (Lightning-TopK): “select” replaces “full sort”, optimizing 80k-level ultra-long inputs
result: GLM-5.2 NVFP4 on Blackwell hits 500+ tok/s/user

Teams running GLM-5.2 should lock post1.

3.3 7/18–7/23 main: speculative decoding keeps getting polished

PR	Content
#31468	DFlash removes per-step host sync in spec decode — CPU leads GPU by a full step (pipeline); branch logic made CUDA-graph-capturable
#31487	reduce prefill CUDA graph padding
#31986	DSpark dense-draft per-layer ctx KV projection stacked into a single big GEMM
#31985	fold draft embedding into the draft graph via `forward_embed`
#31682	DP attention default breakable prefill cuda graph
#31981	DSA draft-extend metadata kernel skips page-table columns beyond KV length
#30272	SM120 DeepSeek V4 flashinfer_mxfp4 MoE + TP2 (consumer Blackwell)
#24013	VLM cross-request ViT encoding batching (multimodal concurrency)
#31825	ModelOpt FP4 path supports NVFP4_AWQ
#31762	fix marlin_nvfp4 `routed_scaling_factor`
#30924	unified per-token-group quant kernel
#30540	new HPC-Ops attention backend
#31109	remove QServe and FBGEMM FP8 quant (old paths migrate to ModelOpt / compressed-tensors)
#32047	fix stale per-token-group quant callers; `evict_from_tree_cache` only evicts the KV gap
#27894 / #31835	NIXL disagg functional tests into CI; PrefillDelayer delays until KV-budget admission

Read #31468, #31986, #31985 together and a very clear main line emerges: eliminate one by one every "ask the CPU every step" in speculative decoding, merge small kernels into big kernels, finally let the whole step into CUDA Graph. What's saved is accelerator idle time, not compute.

3.4 v0.5.16 (7/25): UnifiedRadixTree

UnifiedRadixTree becomes the default prefix cache — unifying previously scattered cache paths (ordinary prefix cache, hierarchical cache, PD-scenario cache) into one tree for hit and eviction logic.

3.5 7/29–7/31

PR #30747 fixes a PP (pipeline parallel) + structured-output crash (Issue #28424). Root cause: the scheduler and constraint logic raced for state concurrently; fix: align the constraint check with the micro-batch boundary.
PR #30440 grpc disaggregated generation.

4. Upgrade Checklist

From a month of traps, a practical checklist:

vLLM 0.24.x → 0.25.x/0.26.x

Go straight to 0.25.1 or higher, don’t stop at 0.25.0 (NVFP4 silent garbage)
Upgrade transformers to v5
Compile environment supports C++20
Check whether you use custom ops (may fall back to MRv1 and lose the gains)
Check whether partial-prefill related old params were removed
From 0.26, check whether startup flags were renamed; confirm FlashInfer is installed for NVFP4
Run an output-correctness regression, don’t just watch throughput

SGLang 0.5.x

Versions below 0.5.15 must upgrade (multimodal path-traversal vulnerability)
Running GLM-5.2 → lock v0.5.15.post1
Used QServe / FBGEMM FP8 → migrate to ModelOpt / compressed-tensors
Want DSpark / DFlash latest → need main, but wait for a stable tag in production
Running PP + structured output together → confirm it includes #30747

5. What the Timeline Tells Us

The two move in tight lockstep: 0.25.0 (7/11) vs 0.5.15 (7/10), 0.26.0 vs 0.5.16 (both 7/25). This means they face the same bottlenecks, not separate wheel-reinvention.
A big release is always followed by a wave of fixes: a patch within a week of 0.25.0, then two weeks of dense edge-config fixes. Waiting 1–2 weeks after a new version before production is the rational choice.
The gap between main and release is large: SGLang’s frontier optimizations (DSpark GEMM stacking, etc.) are all on main; stable tags lag. Want performance → follow main and benchmark yourself; want stability → lock the tag.
The main axis is migrating: early July was “kill sync stalls”; after the same-day late-July release, the axis shifted to speculative decoding + prefix cache + day-one model support + production deployment.

The next post is the most practical one: which models are supported, and who to pick for what scenario.