Qwen3.5-122B-A10B-FP8

Qwen's 122B hybrid MoE (Gated DeltaNet and attention layers, 256 experts, 10B active) reasoning model from the official FP8 checkpoint, text only, at its native 256K context on four Gaudi 2 cards with expert parallelism. Prefix caching is on through the gdn-prefix-cache fix.

Model	Qwen/Qwen3.5-122B-A10B-FP8
Hardware	4x Gaudi 2 (TP 4, expert parallel)
Image	vllm-gaudi v0.26.0 (b059f9c), vLLM 568afb3a, Gaudi SW 1.24.1
Precision	FP8 checkpoint (block 128 x 128, dynamic activations, 127 GB)
Context	262,144 (native)
Concurrency	--max-num-seqs 16
Parsers	tools qwen3_coder, reasoning qwen3
Recipe	recipes/Qwen/Qwen3.5-122B-A10B-FP8/recipe.yaml

Prefix caching

On the stock release, prefix-cache hits on the Gated DeltaNet layers resume from the wrong state, so prefix caching is only correct with gdn-prefix-cache. The fix keeps the GDN layers in their compact per-request slots (VLLM_COMPACT_GDN=1) and stores a state checkpoint at the end of every full KV block in that block's own attention page. A hit restores the checkpoint before the next chunk, and the full KV pool stays available. With caching on, the hybrid KV block is 1,152 tokens and prefill chunks end on block boundaries; prompts shorter than one block never hit. The design and the upstream status are in Prefix caching on Gated DeltaNet hybrids.

Launch

$COOKBOOK is a checkout of this repository inside the container.

export PT_HPU_LAZY_MODE=0
export PT_HPU_ENABLE_LAZY_COLLECTIVES=true
export VLLM_HPU_MOE_GATHER=1
export VLLM_HPU_MOE_GATHER_RATIO=1.0
export VLLM_HPU_MOE_GATHER_VERIFY=0
export VLLM_USE_NAIVE_MAMBA_CACHE_SHARING=1
export VLLM_COMPACT_GDN=1
export PT_HPUGRAPH_DISABLE_TENSOR_CACHE=false
export VLLM_SKIP_WARMUP=true
export VLLM_GRAPH_RESERVED_MEM=0.3
export VLLM_EXPONENTIAL_BUCKETING=True
export VLLM_PROMPT_BS_BUCKET_MIN=1
export VLLM_PROMPT_BS_BUCKET_STEP=1
export VLLM_PROMPT_BS_BUCKET_MAX=16
export VLLM_DECODE_BS_BUCKET_MIN=1
export VLLM_DECODE_BS_BUCKET_STEP=1
export VLLM_DECODE_BS_BUCKET_MAX=16
export VLLM_PROMPT_SEQ_BUCKET_MIN=128
export VLLM_PROMPT_SEQ_BUCKET_STEP=4096
export VLLM_PROMPT_SEQ_BUCKET_MAX=262144
export VLLM_DECODE_BLOCK_BUCKET_MIN=128
export VLLM_DECODE_BLOCK_BUCKET_STEP=1024
export VLLM_DECODE_BLOCK_BUCKET_MAX=16384
export VLLM_ENGINE_ITERATION_TIMEOUT_S=7200
export VLLM_RPC_TIMEOUT=7200000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=1800

# Fixes: the v0.26.0 profile plus this recipe's opt-in fixes, in the pack's order.
# Stops with a non-zero exit code if the image does not match.
python3 "$COOKBOOK/fixes/apply_all.py" --profile v0.26.0 gdn-prefix-cache fp8-moe-gather || exit 1

vllm serve Qwen/Qwen3.5-122B-A10B-FP8 \
  --tensor-parallel-size 4 \
  --enable-expert-parallel \
  --max-model-len 262144 \
  --max-num-seqs 16 \
  --gpu-memory-utilization 0.75 \
  --block-size 128 \
  --max-num-batched-tokens 4096 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --enable-prompt-tokens-details \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --language-model-only

• Compile mode (PT_HPU_LAZY_MODE=0), with the compact GDN state layout that gdn-prefix-cache needs.
• --enable-expert-parallel puts 64 of the 256 experts on each card.
• fp8-moe-gather with VLLM_HPU_MOE_GATHER=1: at low token counts only the routed experts are read. The ratio is per card, so with 64 local experts 1.0 keeps the gather path for up to 8 decode tokens (8 x 8 routed experts); larger steps use the stock MoE op.
• --language-model-only serves the model as text only and skips the vision tower's memory profiling.
• --max-num-seqs 16 and --gpu-memory-utilization 0.75 are the values the deployment ran with; per-stream decode falls as concurrency rises, so 16 rather than 32.
• With prefix caching on, prefill chunks end on 1,152-token blocks: the 4,096-token budget runs as 3 blocks (3,456 tokens) per chunk.

Buckets

No bucket file. Prompt and decode batch buckets 1 to 16, prompt length 128 to 262,144 in steps of 4,096 with exponential bucketing, decode blocks 128 to 16,384 in steps of 1,024.

Required fixes

All from Fixes. The ones marked opt-in are named in the launch command; the rest come from --profile v0.26.0, which also applies the release's other generic fixes (see Using a recipe). The profile also runs seeded-sampling; this recipe does not depend on it (the seed segfault is a lazy-mode bug).

Fix	Why this model needs it
prompt-order	with prefix caching, a request that hits a block another request writes in the same step read it before it was written
sampling-rows-refresh	stale device sampling rows after a batch reorder
structured-output-rows	structured-output steps sample each row with another request's parameters
bad-words-index	bad_words raises IndexError in the engine core
logitsprocs-rows	logit_bias, min_p and min_tokens are silently ignored
async-discard-counter	AttributeError on the first preemption under async scheduling
resumed-prefill-logits	a request resumed after preemption with prefix-cache hits samples again at every recomputed position, so the generated tail repeats
hybrid-decode-pad	when the hybrid KV block is split into 128-token kernel blocks, decode padding writes into a live block of another request
gdn-prefix-cache (opt-in)	prefix-cache hits on the GDN layers resume from a wrong state on the stock release; the fix stores and restores GDN checkpoints in the KV pages and zeroes new blocks and slots. It needs hybrid-decode-pad first
gather-logprobs-int32	the id corruption is lazy-mode only, but the fix is harmless here and keeps one fix set across recipes
prompt-logprobs	qwen3_5_moe returns flat hidden states (also with --language-model-only); the first prompt_logprobs request kills the engine
logprobs-decode-guard	a bad logprob id from any source is replaced instead of stopping the API server
fp8-moe-gather (opt-in)	gathered-expert FP8 MoE, a backport from v0.30.0: faster low-batch decode
rope-cpu-cache	Gaudi fp32 cos/sin are wrong past about 1.5e5, so RoPE tables built on the HPU are wrong from position about 196K
rank-exception-fatal	an exception on a non-zero TP rank stops the engine instead of serving from out-of-step ranks

Verified

On vllm-gaudi v0.26.0, four Gaudi 2 cards, with exactly the launch command above in a container started from the stock image (the fixes applied at start-up). Requests went to the OpenAI-compatible API.

Check	Result
Known-answer prompts (arithmetic, facts, translation), temperature 0	8/8
The same prompts, 16 in flight	16/16
Tool call, tool_choice: auto, sampled at 0.6	5/5 valid get_weather calls with the right argument
Streamed tool call	name and JSON arguments assembled correctly
Tool result fed back	the answer uses the returned values
top_logprobs: 20	20 entries on every token
prompt_logprobs: 5	returned; server healthy afterwards
response_format JSON schema	valid object
logit_bias +100 on one token	takes effect
Prefix-cache hit vs cold (cache_salt), greedy, 8 prompts of 6K to 126K tokens with a needle	identical text and bit-identical logprobs 8/8; cached tokens a multiple of 1,152 on every hit
Needle in a haystack, 3 depths, cold	3/3 at 65K, 131K and 244K tokens
Server health after all checks	healthy, no restarts

KV pool at start-up: 2,885,824 tokens.

Performance

Single runs on the server above (measured as: per-call, one server; times are end to end and include time to first token).

Test	Result
One stream, short prompt, 256 tokens	43.8 tokens/s
16 streams, short prompts, 256 tokens each	315.6 tokens/s aggregate, 19.7 per stream
Cold prompt of 65K tokens (needle)	16 to 25 s to the answer
Cold prompt of 131K tokens (needle)	24 to 33 s to the answer
Cold prompt of 244K tokens (needle)	56 to 82 s to the answer

Known issues

• Loading is slow under a CPU quota: vllm-gaudi sizes its load thread pool from the host's core count, and the per-tensor FP8 rescale on Gaudi 2 is then CPU-throttled. Run without a CPU limit, or cap the pool at the quota.
• Give test requests a real token budget: this model reasons at length.

What not to do

• Do not use lazy mode for this model.
• Do not turn on prefix caching without gdn-prefix-cache, and do not set VLLM_COMPACT_GDN=0 with it.
• Do not copy VLLM_HPU_MOE_GATHER_RATIO to a layout without expert parallelism: the ratio is relative to the experts on one card.