[Klaud Cold] Move golden_al_distribution/ into infx/

.github/codeowner-signoff-verify-prompt.md

@@ -324,7 +324,7 @@ speculative decoding. From the PR diff, identify configs that are BOTH:
   downloads, or config names containing `-mtp` / `eagle`.
 Agentic replay does not reproduce real-world token-by-token traffic, so measured
 acceptance there is not representative. Per the AgentX fairness guidelines
-in `golden_al_distribution/README.md` on the checked-out default branch, such configs
+in `infx/golden_al_distribution/README.md` on the checked-out default branch, such configs
 must instead SIMULATE acceptance at the committed golden acceptance length (AL).
 Verify BOTH:
 - (a) SIMULATED ACCEPTANCE ENABLED. The launch config must pin a simulated/synthetic
@@ -343,7 +343,7 @@ Verify BOTH:
   FAIL if an agentic spec-decode config runs real (unsimulated) acceptance.
   Name the config/script and line.
 - (b) AL VALUE MATCHES THE GOLDEN CURVE. Read the committed golden AL YAML for the
-  model in `golden_al_distribution/` on the default-branch checkout. Examples include
+  model in `infx/golden_al_distribution/` on the default-branch checkout. Examples include
   `qwen3.5_mtp.yaml` and `minimaxm3_eagle3.yaml`. Confirm the pinned AL equals the golden value for that
   model, thinking mode, and the config's `num_speculative_tokens` / MTP level (e.g.
   qwen3.5 thinking_on with 3 speculative tokens -> 3.39). For TRT-LLM configs, compare

AGENTS.md

@@ -75,7 +75,7 @@ check_env_vars IS_MULTINODE MODEL_NAME PRECISION
 
 ## SRT Slurm synthetic acceptance
 
-- **Do not hard-code synthetic acceptance lengths in SRT recipes, master configs, or launchers.** InferenceX automatically selects the measured value from [`golden_al_distribution/`](golden_al_distribution/) for speculative AgentX throughput runs. Do not add manual `SYNTHETIC_ACCEPTANCE_LENGTH`, vLLM `synthetic_acceptance_length`, SGLang `SGLANG_SIMULATE_ACC_LEN`, or TRT-LLM `TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS` settings.
+- **Do not hard-code synthetic acceptance lengths in SRT recipes, master configs, or launchers.** InferenceX automatically selects the measured value from [`infx/golden_al_distribution/`](infx/golden_al_distribution/) for speculative AgentX throughput runs. Do not add manual `SYNTHETIC_ACCEPTANCE_LENGTH`, vLLM `synthetic_acceptance_length`, SGLang `SGLANG_SIMULATE_ACC_LEN`, or TRT-LLM `TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS` settings.
 - Submit recipes through [`apply_srt_recipe`](runners/slurm_utils.sh). Its [`infx/srt_slurm` connector](infx/srt_slurm/synthetic_acceptance.py) applies native SRT `--set` / `--unset` overrides; calling upstream `srtctl` directly does not perform InferenceX's automatic selection.
 - Keep the actual speculative method, draft model, draft-token count, and relevant sampling settings explicit in the recipe. The connector combines the generation role's settings (decode, otherwise aggregated), after caller overrides, with `MODEL_PREFIX` and `THINKING_MODE` to select the golden curve. For Kimi DSpark, explicitly set `draft_sample_method` to `greedy` or `probabilistic`.
 - Eval-only and non-AgentX runs use real verification; the connector removes stale synthetic settings. Non-speculative roles do not receive simulation settings. `RUN_EVAL` does not disable simulation for the throughput portion.

benchmarks/multi_node/amd_utils/models.yaml

@@ -87,5 +87,5 @@ DeepSeek-V4-Pro-AgentX: &DeepSeek-V4-Pro-AgentX
 # measured clean across c4-c256; the EAGLE 3-1-4 arm remains available through
 # spec-decoding: mtp. Synthetic acceptance is checkpoint-specific in
 # server_sglang.sh and does not depend on which of the two runs: thinking-on,
-# draft lengths 1-8 are wired from golden_al_distribution/dsv4-pro-0813-dspark.yaml.
+# draft lengths 1-8 are wired from infx/golden_al_distribution/dsv4-pro-0813-dspark.yaml.
 DeepSeek-V4-Pro-0813-AgentX: *DeepSeek-V4-Pro-AgentX

benchmarks/multi_node/amd_utils/server_sglang.sh

@@ -1418,15 +1418,15 @@ else
     # Agentic trace replay doesn't reproduce real token-by-token traffic, so
     # measured MTP/EAGLE acceptance there isn't representative (PR #2309
     # review: https://github.com/SemiAnalysisAI/InferenceX/pull/2309#pullrequestreview-4778348624).
-    # Per the AgentX fairness guidelines (golden_al_distribution/README.md),
+    # Per the AgentX fairness guidelines (infx/golden_al_distribution/README.md),
     # agentic throughput benchmarks simulate acceptance at the model's
     # committed golden AL instead of measuring real (non-representative)
     # acceptance. Eval runs (RUN_EVAL / EVAL_ONLY) need real acceptance so
     # GSM8K scores reflect actual MTP behavior. AgentX uses one golden curve
     # per checkpoint, thinking mode, and draft length, including when the
     # supported PD draft implementation differs from the calibration engine.
     # Sources (thinking_on): dsv4_mtp.yaml for the original checkpoint and
-    # golden_al_distribution/dsv4-pro-0813-dspark.yaml for Pro-0813.
+    # infx/golden_al_distribution/dsv4-pro-0813-dspark.yaml for Pro-0813.
     DECODE_SIM_ACC_ENV=""
     if [[ "$DECODE_MTP_SIZE" -gt 0 ]] && { [[ "${IS_AGENTIC}" == "1" ]] || [[ "${IS_AGENTIC:-}" == "true" ]]; }; then
         if [[ "${EVAL_ONLY}" == "true" ]] || [[ "${RUN_EVAL}" == "true" ]]; then
@@ -1453,7 +1453,7 @@ else
             if [[ -n "$DSV4_GOLDEN_AL" ]]; then
                 DECODE_SIM_ACC_ENV="SGLANG_SIMULATE_ACC_LEN=${DSV4_GOLDEN_AL} SGLANG_SIMULATE_ACC_METHOD=match-expected SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token"
             else
-                echo "WARNING: agentic spec-decoding run (model=${MODEL_NAME}, algorithm=${SPEC_DECODING:-mtp}, DECODE_MTP_SIZE=${DECODE_MTP_SIZE}) has no golden AL wired in server_sglang.sh -- falling back to real (unsimulated, non-representative) acceptance. Add a case in server_sglang.sh and golden_al_distribution/ before shipping this arm. See golden_al_distribution/README.md." >&2
+                echo "WARNING: agentic spec-decoding run (model=${MODEL_NAME}, algorithm=${SPEC_DECODING:-mtp}, DECODE_MTP_SIZE=${DECODE_MTP_SIZE}) has no golden AL wired in server_sglang.sh -- falling back to real (unsimulated, non-representative) acceptance. Add a case in server_sglang.sh and infx/golden_al_distribution/ before shipping this arm. See infx/golden_al_distribution/README.md." >&2
             fi
         fi
     fi

benchmarks/multi_node/amd_utils/server_tilert.sh

@@ -241,7 +241,7 @@ start_decode() {
     local extra=( ${TILERT_DECODE_EXTRA_FLAGS} )
     if [[ "$SPEC_DECODING" == "mtp" && "$EVAL_ONLY" != "true" && "$RUN_EVAL" != "true" ]]; then
         check_env_vars MODEL_PREFIX THINKING_MODE
-        local curve="${WS_PATH%/benchmarks/*}/golden_al_distribution/${MODEL_PREFIX}_mtp.yaml"
+        local curve="${WS_PATH%/benchmarks/*}/infx/golden_al_distribution/${MODEL_PREFIX}_mtp.yaml"
         TILERT_SIMULATE_ACC_LEN="$("$PY" - "$curve" "$THINKING_MODE" "$DECODE_MTP_SIZE" <<'PYEOF'
 import sys, yaml
 path, thinking, tokens = sys.argv[1], sys.argv[2], int(sys.argv[3])

benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml

@@ -125,7 +125,7 @@ base:
 # One TP8 worker serves prefill and decode on a single node. The sweep matrix
 # supplies the concurrency list; srt_agentic.sh replays every point against
 # this one server. Acceptance is pinned to the golden thinking-on AL for three
-# speculative tokens (golden_al_distribution/glm5.2_mtp.yaml).
+# speculative tokens (infx/golden_al_distribution/glm5.2_mtp.yaml).
 override_c1:
   name: agg-b200-tp8-c1-mtp
   roles:
@@ -139,7 +139,7 @@ override_c1:
 # One TP8 worker serves prefill and decode on a single node. The sweep matrix
 # supplies the concurrency list; srt_agentic.sh replays every point against
 # this one server. Acceptance is pinned to the golden thinking-on AL for three
-# speculative tokens (golden_al_distribution/glm5.2_mtp.yaml).
+# speculative tokens (infx/golden_al_distribution/glm5.2_mtp.yaml).
 override_c4:
   name: agg-b200-tp8-c4-mtp
   roles:
@@ -153,7 +153,7 @@ override_c4:
 # One TP8 worker serves prefill and decode on a single node. The sweep matrix
 # supplies the concurrency list; srt_agentic.sh replays every point against
 # this one server. Acceptance is pinned to the golden thinking-on AL for three
-# speculative tokens (golden_al_distribution/glm5.2_mtp.yaml).
+# speculative tokens (infx/golden_al_distribution/glm5.2_mtp.yaml).
 override_c8:
   name: agg-b200-tp8-c8-mtp
   roles:

benchmarks/single_node/speedbench/dsv4dspark_fp4_b300_vllm.sh

@@ -39,10 +39,10 @@ MAX_NUM_SEQS="64"
 # Reserve device memory for KV cache and speculative verification.
 GPU_MEM_UTIL="0.90"
 TEMPERATURE="1.0"
-# MUST match the golden config: golden_al_distribution/dsv4_mtp.yaml was measured
+# MUST match the golden config: infx/golden_al_distribution/dsv4_mtp.yaml was measured
 # with reasoning_effort=high.
 # The published recipe uses greedy; probabilistic won at every level on Kimi-K3
-# (golden_al_distribution/kimik3_dspark*.yaml). vLLM accepts exactly these two values
+# (infx/golden_al_distribution/kimik3_dspark*.yaml). vLLM accepts exactly these two values
 # (vllm/config/speculative.py: DraftSampleMethod).
 case "$DRAFT_SAMPLE_METHOD" in
     greedy|probabilistic) ;;

configs/amd-master.yaml

@@ -654,7 +654,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in
-# golden_al_distribution/kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml
+# infx/golden_al_distribution/kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml
 # at both draft lengths used here: 7 draft tokens -> AL 3.84 at concurrency 1 and 4,
 # 3 draft tokens -> AL 3.00 at concurrency 14 and 16.
 # Companion to kimik3-fp4-mi355x-vllm-agentic-mtp: same checkpoint, same runner.
@@ -1411,7 +1411,7 @@ dsv41flash-fp4-mi355x-vllm-agentic-dspark:
 # and draft length (docs/PR_REVIEW_CHECKLIST.md), and a submission may not
 # substitute its own target. Nothing about that target is written here:
 # AGENTS.md forbids hard-coding an acceptance length in a master config, so
-# server_tilert.sh reads golden_al_distribution/glm5.3_mtp.yaml at launch and
+# server_tilert.sh reads infx/golden_al_distribution/glm5.3_mtp.yaml at launch and
 # fails the run if the curve or the draft length is missing.
 # The curve is consumed in the same units as every other framework here, and
 # AgentX replays run with thinking on.

configs/nvidia-master.yaml

@@ -1502,7 +1502,7 @@ dsr1-fp8-b300-sglang-mtp:
 kimik3-fp4-b300-vllm-agentic-dspark:
   # TP8 x DCP8 with Mooncake as the external KV tier. The recipe drafts with
   # DSpark level 7 at conc <= 8 and stops above it; synthetic acceptance is
-  # pinned to the committed golden AL 3.84 (golden_al_distribution/
+  # pinned to the committed golden AL 3.84 (infx/golden_al_distribution/
   # kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml),
   # and EVAL_ONLY switches to real block verification.
   #
@@ -5177,7 +5177,7 @@ qwen3.5-fp8-h100-sglang-agentic:
 # qwen3.5-fp8-h100-sglang-agentic: same TP8/EP8 GPU-resident and HiCache arms,
 # plus SGLang EAGLE MTP (num-steps 3, eagle-topk 1, 4 draft tokens = 3
 # speculative tokens) with simulated acceptance pinned to the golden AL 3.39
-# (golden_al_distribution/qwen3.5_mtp.yaml, thinking_on, K=3) -- the same value
+# (infx/golden_al_distribution/qwen3.5_mtp.yaml, thinking_on, K=3) -- the same value
 # the GB300 Qwen3.5 AgentX srt-slurm recipes use. Image moves to
 # lmsysorg/sglang:v0.5.16-cu130 because SGLANG_SIMULATE_ACC_TOKEN_MODE only
 # exists from v0.5.16; v0.5.14 is the version the fixed-seq-len H100 MTP entry
@@ -5187,7 +5187,7 @@ qwen3.5-fp8-h100-sglang-agentic:
 # policy that new agentic arms enable speculative decoding rather than running a
 # separate STP baseline (docs/MODELS.md). SGLang EAGLE MTP (num-steps 3, eagle-topk 1,
 # 4 draft tokens = 3 speculative tokens) with simulated acceptance pinned to the
-# golden AL 3.39 (golden_al_distribution/qwen3.5_mtp.yaml, thinking_on, K=3).
+# golden AL 3.39 (infx/golden_al_distribution/qwen3.5_mtp.yaml, thinking_on, K=3).
 # Image is lmsysorg/sglang:v0.5.16-cu130: SGLANG_SIMULATE_ACC_TOKEN_MODE only
 # exists from v0.5.16, and v0.5.14 is what the fixed-seq-len H200 MTP entry runs.
 # Search space is the H100 AgentX shape widened for H200's 141 GB: the
@@ -7020,7 +7020,7 @@ glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-agg:
 # AgentX speculative-decoding policy in docs/MODELS.md. SGLang EAGLE runs off
 # GLM-5.2's built-in nextn head (num-steps 3,
 # eagle-topk 1, 4 draft tokens = 3 speculative tokens), with acceptance pinned
-# to the golden AL 2.99 (golden_al_distribution/glm5.2_mtp.yaml, thinking_on,
+# to the golden AL 2.99 (infx/golden_al_distribution/glm5.2_mtp.yaml, thinking_on,
 # K=3) through SGLANG_SIMULATE_ACC_*. The pinned nightly includes FlashInfer
 # 0.6.18's BF16 TRTLLM MoE allocation fix for small-batch Blackwell execution
 # in GLM-5.2's unquantized EAGLE draft head.
@@ -7048,7 +7048,7 @@ glm5.2-fp4-b300-sglang-agentic-mtp:
 # glm5.2-fp8-b200-sglang-agentic-mtp (#2863). Same spec-decode-only shape per
 # the AgentX policy (docs/MODELS.md): SGLang EAGLE off GLM-5.2's built-in nextn head
 # (num-steps 3, eagle-topk 1, 4 draft tokens = 3 speculative tokens) with
-# acceptance pinned to the golden AL 2.99 (golden_al_distribution/glm5.2_mtp.yaml,
+# acceptance pinned to the golden AL 2.99 (infx/golden_al_distribution/glm5.2_mtp.yaml,
 # thinking_on, K=3), which was measured on glm-5.2-fp8, i.e. on this checkpoint.
 # Pinned to the 2026-09-08 cu13 dev nightly (the first that carries sgl-project/sglang#38318, the guard for the EAGLE DSA fp8 read-door crash seen on the 2026-09-07 build), the same tag the B200/B300 GLM-5.2
 # and Qwen3.5 SGLang AgentX recipes moved to that day. Runs on cluster:b300-dsxe
@@ -7078,7 +7078,7 @@ glm5.2-fp8-b300-sglang-agentic-mtp:
 # policy that agentic arms enable speculative decoding rather than running a
 # separate STP baseline (docs/MODELS.md). SGLang EAGLE off GLM-5.2's built-in nextn
 # head (num-steps 3, eagle-topk 1, 4 draft tokens = 3 speculative tokens) with
-# acceptance pinned to the golden AL 2.99 (golden_al_distribution/glm5.2_mtp.yaml,
+# acceptance pinned to the golden AL 2.99 (infx/golden_al_distribution/glm5.2_mtp.yaml,
 # thinking_on, K=3) through SGLANG_SIMULATE_ACC_*. SGLang v0.5.16 is the first
 # release that reads SGLANG_SIMULATE_ACC_TOKEN_MODE. The pinned 2026-09-01
 # nightly includes FlashInfer 0.6.18's BF16 TRTLLM MoE allocation fix, which is
@@ -7107,7 +7107,7 @@ glm5.2-fp4-b200-sglang-agentic-mtp:
 # glm5.2-fp4-b200-sglang-agentic-mtp. Same spec-decode-only shape per the AgentX
 # policy (docs/MODELS.md): SGLang EAGLE off GLM-5.2's built-in nextn head (num-steps
 # 3, eagle-topk 1, 4 draft tokens = 3 speculative tokens) with acceptance pinned
-# to the golden AL 2.99 (golden_al_distribution/glm5.2_mtp.yaml, thinking_on,
+# to the golden AL 2.99 (infx/golden_al_distribution/glm5.2_mtp.yaml, thinking_on,
 # K=3) through SGLANG_SIMULATE_ACC_*. That curve was measured on glm-5.2-fp8,
 # i.e. on this checkpoint. Pinned to the 2026-09-08 cu13 dev nightly (the first that carries sgl-project/sglang#38318, the guard for the EAGLE DSA fp8 read-door crash seen on the 2026-09-07 build), the same
 # tag the B200 Qwen3.5 FP8/FP4 SGLang AgentX recipes moved to that day.
@@ -7135,7 +7135,7 @@ glm5.2-fp8-b200-sglang-agentic-mtp:
 
 # GLM-5.2 NVFP4 B200 AgentX on Dynamo + SGLang. EAGLE uses the model's
 # built-in nextn head, with acceptance pinned to the golden thinking-on AL in
-# golden_al_distribution/glm5.2_mtp.yaml. Both arms use HiCache write-back
+# infx/golden_al_distribution/glm5.2_mtp.yaml. Both arms use HiCache write-back
 # DRAM offload for the agentic working set.
 glm5.2-fp4-b200-dynamo-sglang-agentic-agg:
   image: lmsysorg/sglang:nightly-dev-20260910-00840301

docs/MODELS.md

@@ -34,7 +34,7 @@ The A/B retirements below concern standalone non-spec-decode baselines maintaine
 
 **Status: baseline retirement is not yet enacted.** Retire a redundant baseline only once replacement coverage exists for the affected model, hardware, and engine. Keep non-spec-decode configurations that contribute to the Pareto frontier; the existence of a spec-decode counterpart alone is not a reason to remove them.
 
-**Going forward we no longer maintain separate non-spec-decode and spec-decode tracks solely as an A/B comparison.** The non-spec-decode arm existed as a neutral baseline back when acceptance length wasn't standardized. That is now solved. [`golden_al_distribution/`](../golden_al_distribution/) commits one golden acceptance-length curve per model, thinking mode, and draft length, measured on the SPEED-Bench `coding` category. When speculative decoding is enabled, AgentX pins submissions to that curve through synthetic acceptance (vLLM `synthetic_acceptance_length`, SGLang `SGLANG_SIMULATE_ACC_LEN`, TensorRT-LLM `TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS`, etc). With a fair, engine-independent acceptance target in place, spec-decode results are directly comparable on their own and a dedicated non-spec-decode baseline is redundant. New models, including Kimi-K3, do not require that separate baseline from day 0.
+**Going forward we no longer maintain separate non-spec-decode and spec-decode tracks solely as an A/B comparison.** The non-spec-decode arm existed as a neutral baseline back when acceptance length wasn't standardized. That is now solved. [`infx/golden_al_distribution/`](../infx/golden_al_distribution/) commits one golden acceptance-length curve per model, thinking mode, and draft length, measured on the SPEED-Bench `coding` category. When speculative decoding is enabled, AgentX pins submissions to that curve through synthetic acceptance (vLLM `synthetic_acceptance_length`, SGLang `SGLANG_SIMULATE_ACC_LEN`, TensorRT-LLM `TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS`, etc). With a fair, engine-independent acceptance target in place, spec-decode results are directly comparable on their own and a dedicated non-spec-decode baseline is redundant. New models, including Kimi-K3, do not require that separate baseline from day 0.
 
 **Publish the best Pareto points, whether speculative decoding is enabled or disabled.** Recipes may disable MTP, EAGLE/EAGLE3, DSpark, or another draft method when doing so produces a better operating point, for example at high throughput. Valid non-spec-decode results remain eligible for publication under the same [North-star Pareto policy](#north-star-pareto-policy). A frontier may therefore contain both spec-decode and non-spec-decode points.