From 0fe7550254674d843d645c4e01c5ceaf6bf87e81 Mon Sep 17 00:00:00 2001 From: Benjamin Chislett Date: Fri, 15 May 2026 10:29:45 -0400 Subject: [PATCH] [Bugfix] DFlash FP8 KV-Cache (#42692) Signed-off-by: Benjamin Chislett --- vllm/model_executor/models/qwen3_dflash.py | 4 +++- vllm/v1/spec_decode/llm_base_proposer.py | 2 +- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/models/qwen3_dflash.py b/vllm/model_executor/models/qwen3_dflash.py index dcc23bc409..231ed646e0 100644 --- a/vllm/model_executor/models/qwen3_dflash.py +++ b/vllm/model_executor/models/qwen3_dflash.py @@ -247,8 +247,10 @@ class DFlashQwen3Model(nn.Module): [ DFlashQwen3DecoderLayer( current_vllm_config, - prefix=maybe_prefix(prefix, f"layers.{layer_idx + start_layer_id}"), config=self.config, + cache_config=current_vllm_config.cache_config, + quant_config=self.quant_config, + prefix=maybe_prefix(prefix, f"layers.{layer_idx + start_layer_id}"), ) for layer_idx in range(self.config.num_hidden_layers) ] diff --git a/vllm/v1/spec_decode/llm_base_proposer.py b/vllm/v1/spec_decode/llm_base_proposer.py index 8c9f2f7bae..dc50867329 100644 --- a/vllm/v1/spec_decode/llm_base_proposer.py +++ b/vllm/v1/spec_decode/llm_base_proposer.py @@ -116,7 +116,7 @@ class SpecDecodeBaseProposer: self.max_batch_size = vllm_config.scheduler_config.max_num_seqs self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens - self.token_arange_np = np.arange(self.max_num_tokens) + self.token_arange_np = np.arange(self.max_num_tokens, dtype=np.int32) # Can be specialized by methods like DFlash to reduce the limit self.max_query_tokens = self.max_num_tokens