fix: swap reasoning-predictor to Qwen3-32B-GPTQ-Int4, 131072 context (bnb-4bit decode too slow, GPTQ is Volta-native)

k8s/apps/llm-serving/reasoning.yaml

@@ -12,25 +12,40 @@ spec:
   predictor:
     containers:
     - args:
-      # DeepSeek-R1-Distill-32B retired: tool_choice="auto" (what pi sends)
-      # hit a documented vLLM/R1-family conflict -- the model narrated fake
-      # tool_calls in its <think> block instead of emitting real ones,
-      # regardless of parser. Tried swapping to a Kimi-distilled Qwen3.6
-      # MoE checkpoint and an AWQ-quantized Qwen3-30B-A3B first -- both
-      # failed on real, separate blockers (unrecognized model_type; then
-      # marlin INT4 kernels needing compute capability 80+, but worker-1's
-      # GPU is sm70/V100). Landed on dense Qwen3-32B instead: native Qwen3
-      # tool-call format (no narration bug), bnb-4bit works fine on sm70
-      # (proven by the old DeepSeek config already), and no MoE
-      # arch/quantization risk this time.
-      - --model=unsloth/Qwen3-32B-bnb-4bit
+      # bnb-4bit retired: no int4 tensor cores on sm70/V100, dequant-then-
+      # matmul is two slow kernel launches instead of one fused int4 GEMM,
+      # decode crawled at 2.5-10 tok/s regardless of TP/PP. Switched to
+      # JunHowie/Qwen3-32B-GPTQ-Int4 -- same dense Qwen3-32B weights, same
+      # hermes/qwen3 parser stack (no narration-bug risk, same as before),
+      # only the quant format changes. Plain (non-Marlin) GPTQ kernel is
+      # confirmed Volta-compatible; Marlin needs sm80+ and vLLM would try
+      # to auto-upgrade to it, so --quantization is pinned explicitly to
+      # `gptq` to force the plain kernel. Verified checkpoint size: 19.34GB
+      # (summed from the real safetensors index, not bits-per-param math).
+      # max-model-len=131072 is Qwen3-32B's real ceiling (config.json YaRN:
+      # factor=4.0, original_max_position_embeddings=32768) -- 200k was
+      # asked for but exceeds this architecturally regardless of VRAM.
+      # KV cache math: 256KB/token total (64 layers, 8 KV heads, 128
+      # head_dim, fp16), PP=2 splits both weights and KV load ~evenly, so
+      # each GPU carries ~9.67GB weights + ~128KB/token KV. At
+      # gpu-memory-utilization=0.90 (28.8GB/GPU usable), that leaves
+      # ~19.1GB/GPU for KV cache -> ~156k tokens/GPU capacity, comfortably
+      # above the 131072 target with room to spare -- the old
+      # OffloadingConnector CPU-DRAM spillover (tuned for the previous
+      # model's much smaller 16384 context) is no longer needed and is
+      # dropped. Staying on PP=2 and vLLM 0.11.0 (no version bump needed,
+      # this checkpoint only requires vllm>=0.9.2) -- plain GPTQ has no
+      # TP>1 restriction unlike bnb, so tensor-parallel-size=2 is worth
+      # trying later, but not risking a parallelism-strategy change in the
+      # same rollout as the quant+context-length change.
+      - --model=JunHowie/Qwen3-32B-GPTQ-Int4
       - --served-model-name=reasoning
-      - --quantization=bitsandbytes
+      - --quantization=gptq
       - --dtype=float16
       - --kv-cache-dtype=auto
       - --tensor-parallel-size=1
       - --pipeline-parallel-size=2
-      - --max-model-len=16384
+      - --max-model-len=131072
       - --gpu-memory-utilization=0.90
       - --max-num-seqs=4
       - --enable-chunked-prefill
@@ -42,27 +57,6 @@ spec:
       # Qwen3 models -- native chat template support, not narrated text.
       - --enable-auto-tool-choice
       - --tool-call-parser=hermes
-      # vLLM 0.11.0's native OffloadingConnector -- spills KV cache blocks
-      # to CPU DRAM instead of discarding them on preemption (max-num-seqs=4
-      # + max-model-len=16384 means concurrent long sequences compete for
-      # the same GPU KV space). No extra dependency, built into vLLM core.
-      # num_cpu_blocks=2000 hung the pod at startup on the old model (2000 x
-      # ~32MB/block blew well past the pod's memory limit). num_cpu_blocks=32
-      # was the safe-recovery value after that -- only ~1GB of real DRAM
-      # (32 blocks x 128 tokens x 256KB/token-across-all-64-layers, fp16),
-      # basically a token-count safety valve, not real offload capacity.
-      # This model: 64 layers, 8 KV heads x 128 head_dim, fp16 -> ~256KB of
-      # KV per token across all layers -> ~32MB per 128-token block.
-      # num_cpu_blocks=256 -> ~8GB of actual DRAM offload (32,768 tokens),
-      # comfortably under the pod's 36Gi limit alongside the ~20GB bnb-4bit
-      # weights. Watch real host memory on boot before raising further --
-      # block_size=128 tokens matches vLLM's own example.
-      # Note: 0.11.0 ships the original (fragmented, small-transfer-block)
-      # version of this connector -- 0.12.0 consolidates KV data into one
-      # contiguous block per request and is reported an order of magnitude
-      # faster for this specific feature, so this is a real but not yet
-      # optimal implementation until the image gets bumped.
-      - --kv-transfer-config={"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"num_cpu_blocks":256,"block_size":128}}
       - --host=0.0.0.0
       - --port=8080
       env: