All issues

Search results

[Bug]: Default memory config cannot boot a 256K-vocab model on v5e-1: the [512]-bucket structured_decode program has no room left to load

bugSomething isn't workingSomething isn't working

Status: Open.

#3360 In vllm-project/tpu-inference;

· jerviscz opened on Aug 9, 2026

[Bug]: TPU GDN batched decode hangs when padded request metadata is narrower than decode_tile_size

bugSomething isn't workingSomething isn't working

Status: Open.

#3356 In vllm-project/tpu-inference;

· zook8 opened on Aug 9, 2026

[Feature]: Fully support vllm tpu inference for Qwen3-VL

Status: Open.

#3350 In vllm-project/tpu-inference;

· capri-xiyue opened on Aug 6, 2026

[Bug]: tpu_raiden dependency is missing in v0.25.0

bugSomething isn't workingSomething isn't working

Status: Open.

#3285 In vllm-project/tpu-inference;

· zxxia opened on Jul 28, 2026

[Bug]: Gemma 4 E2B QAT checkpoints fail to load: k_norm required for KV-shared layers (num_kv_shared_layers ignored by loader)

Status: Open.

#3225 In vllm-project/tpu-inference;

· xbill9 opened on Jul 21, 2026

[Feature] Make Vit Encoder CUDA graph feature work with Qwen3-VL moe

Status: Open.

#3164 In vllm-project/tpu-inference;

· lk-chen opened on Jul 13, 2026

[Performance]: unconditional fp32 logits convert + aliasing copy costs ~12% of the greedy decode step (Gemma 2B, v5e, batch 512)

Status: Open.

#3127 In vllm-project/tpu-inference;

· RubensZimbres opened on Jul 9, 2026

[Bug]: kv_cache_dtype=fp8 silently allocates bf16 on v5e (RPA v3) while logging that fp8 is active

bugSomething isn't workingSomething isn't working

Status: Open.

#3126 In vllm-project/tpu-inference;

· RubensZimbres opened on Jul 9, 2026

[Bug]: BATCHED_RPA_KERNEL does not work with speculative decoding

bugSomething isn't workingSomething isn't working

Status: Open.

#3095 In vllm-project/tpu-inference;

· amanseervi opened on Jul 7, 2026

DeepSeek-V4 dp-attention: per-request seq_len capped at ~num_gpu_blocks/attn_dp_size due to compressor state cache block_size=1

Status: Open.

#3081 In vllm-project/tpu-inference;

· zorrofox opened on Jul 6, 2026

general_device_put OOMs under native jax.distributed multi-host when target sharding is replicated across processes

Status: Open.

#3077 In vllm-project/tpu-inference;

· zorrofox opened on Jul 5, 2026

TPUOffloadConnector: reproducible TPU CoreHalt (E0200) in D2H save path at 8192-token prefill (stable at 2048)

Status: Open.

#3076 In vllm-project/tpu-inference;

· lokic233 opened on Jul 5, 2026