feat: add DeepSeek V4.1 FPM identity and Slurm collection

THIRD_PARTY_NOTICES.md

@@ -67,6 +67,19 @@ This material is licensed under the Apache License 2.0. The upstream license
 at the identified revision is available at:
 https://github.com/ai-dynamo/aiconfigurator/blob/915f590680d8a79fe9c39f6f3a9ff13bc267fcce/LICENSE
 
+## Dynamo V4.1 FPM collection adapter
+
+`collector/fpm_forward/runtime/dsv41/dsv41_scheduler.py` is modified code
+adapted from `components/src/dynamo/vllm/instrumented_scheduler.py` in
+https://github.com/ai-dynamo/dynamo/tree/54960177085413259859c88bd34ed0734d4c2ea9.
+It adds bounded same-request real-KV collection while preserving the native
+benchmark and FPM contracts. Copyright (c) 2025-2026 NVIDIA CORPORATION &
+AFFILIATES. All rights reserved. Licensed under Apache-2.0; the upstream
+license is preserved in the adapter's adjacent `LICENSE`. The adjacent README
+records the inspected vLLM API revision and immutable runtime image/source
+hashes. vLLM implementation files are not vendored. The text fixture and
+lifecycle tests are original work for this change, with no external corpus.
+
 ## NVIDIA AIConfigurator speculative decoding
 
 The speculation SDK, compatibility exports, CLI/task integration, attention and whole-forward FPM operation changes, native bindings, and their tests are adapted and modified from AIConfigurator PR #1563, pinned at commit `6290c161a354da5250c391bd43372b2e9c6f4a51`. Original paths are under `aic-core/src/aiconfigurator_core/sdk/`, `src/aiconfigurator/`, `aic-core/rust/aiconfigurator-core/`, `aic-core/rust/tests/public-api/`, and `tests/`.
@@ -535,7 +548,7 @@ SOFTWARE.
 
 ## Meta Muse Glimmer model configuration
 
-`src/aiconfigurator_core/model_configs/meta-models--Muse-Glimmer-30B_config.json`
+`src/aisimulate_core/model_configs/meta-models--Muse-Glimmer-30B_config.json`
 is an unmodified copy of `config.json` from the Meta Muse Glimmer model
 repository at immutable revision
 `f84ecc3a0ea984a4c04542a84269e3d065350a6e`:
@@ -557,8 +570,8 @@ named Qwen model repositories at the immutable revisions shown:
 
 | Packaged file | Upstream revision |
 | --- | --- |
-| `aiconfigurator_core/model_configs/Qwen--Qwen3.8-2.4T-A95B_config.json` | `Qwen/Qwen3.8-2.4T-A95B@207bd685a7e3696cfaff12ded7c6a7ea0f88c996` |
-| `aiconfigurator_core/model_configs/Qwen--Qwen3.8-2.4T-A95B-FP8_config.json` | `Qwen/Qwen3.8-2.4T-A95B-FP8@d2dc35658bcf77e66643428cb52e774cc3b5bd29` |
+| `aisimulate_core/model_configs/Qwen--Qwen3.8-2.4T-A95B_config.json` | `Qwen/Qwen3.8-2.4T-A95B@207bd685a7e3696cfaff12ded7c6a7ea0f88c996` |
+| `aisimulate_core/model_configs/Qwen--Qwen3.8-2.4T-A95B-FP8_config.json` | `Qwen/Qwen3.8-2.4T-A95B-FP8@d2dc35658bcf77e66643428cb52e774cc3b5bd29` |
 
 Upstream repositories:
 https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B and
@@ -592,10 +605,10 @@ For any questions regarding this license, please contact model-business@notice.q
 The following bundled model configs are modified copies of Meta Llama 4
 checkpoint configuration files:
 
-- `src/aiconfigurator_core/model_configs/meta-llama--Llama-4-Scout-17B-16E-Instruct_config.json`
+- `src/aisimulate_core/model_configs/meta-llama--Llama-4-Scout-17B-16E-Instruct_config.json`
   from revision `92f3b1597a195b523d8d9e5700e57e4fbb8f20d3`:
   https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3/config.json
-- `src/aiconfigurator_core/model_configs/meta-llama--Llama-4-Maverick-17B-128E-Instruct_config.json`
+- `src/aisimulate_core/model_configs/meta-llama--Llama-4-Maverick-17B-128E-Instruct_config.json`
   from revision `73d14711bcc77c16df3470856949c3764056b617`:
   https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct/blob/73d14711bcc77c16df3470856949c3764056b617/config.json
 

crates/core/parity_tests/perfmodel/test_engine_step_parity.py

@@ -2553,7 +2553,9 @@ def test_fpm_spec_tags(self, fpm_systems_root, monkeypatch):
         ctx_op = spec["context_ops"][0]["FpmForward"]
         assert ctx_op["phase"] == "prefill"
         assert spec["generation_ops"][0]["FpmForward"]["phase"] == "decode"
-        assert len(ctx_op["match_identity"]) == 15
+        assert len(ctx_op["match_identity"]) == 19
+        assert ctx_op["match_identity"][-4:] == ["", "full", "none", "text"]
+        assert spec["generation_ops"][0]["FpmForward"]["match_identity"] == ctx_op["match_identity"]
         assert ctx_op["sol_ops"], "sol_ops must carry the original granular list"
 
     @pytest.mark.parametrize(

crates/core/src/engine/common/perf_model.rs

@@ -42,10 +42,12 @@ impl std::fmt::Debug for PerfModel {
 }
 
 impl PerfModel {
-    pub(crate) fn prefill_batch_validation_can_fail(&self) -> bool {
+    // External prediction and duration conversion can fail even when geometry
+    // validation accepts every batch. Keep admission transactional for all providers.
+    pub(crate) fn prefill_pass_can_fail(&self) -> bool {
         match self {
             Self::Polynomial => false,
-            Self::External { timing } => timing.prefill_batch_validation_can_fail(),
+            Self::External { .. } => true,
         }
     }
 

crates/core/src/engine/scheduler/sglang/core.rs

@@ -778,11 +778,11 @@ impl SglangCore {
                 }
             }
         }
-        // Only providers with fallible geometry validation need to preserve the
-        // admission state. Normal polynomial and unrestricted AIC passes avoid
-        // copying radix metadata. Lease checkpoints never become independent owners.
+        // Preserve admission around validation, prediction and duration conversion.
+        // The built-in polynomial remains infallible; external providers may fail
+        // after accepting geometry. Lease checkpoints are not independent owners.
         let admission_checkpoint = (!self.waiting.is_empty()
-            && self.config.perf_model.prefill_batch_validation_can_fail())
+            && self.config.perf_model.prefill_pass_can_fail())
         .then(|| {
             let waiting = self
                 .waiting

crates/core/src/engine/scheduler/sglang/tests.rs

@@ -2658,6 +2658,10 @@ mod admission_validation_rollback {
     }
 
     impl crate::engine::TimingModel for FallibleTiming {
+        fn prefill_batch_validation_can_fail(&self) -> bool {
+            !self.fail_in_prediction
+        }
+
         fn validate_prefill_batch(&self, _: &[(usize, usize)]) -> anyhow::Result<()> {
             anyhow::ensure!(
                 self.fail_in_prediction || !self.fail.load(Ordering::Relaxed),

crates/core/src/engine/timing.rs

@@ -447,9 +447,10 @@ pub enum TimingModelConfig {
 /// Implementations may call AIC, interpolate profiler data, or use another
 /// provider without adding that dependency to `aisimulate-core`.
 pub trait TimingModel: Send + Sync {
-    /// Whether admission needs a checkpoint around `validate_prefill_batch`.
-    /// Existing custom providers default to the safe, fallible contract. Providers
-    /// opting out must accept every batch geometry in that validation hook.
+    /// Whether `validate_prefill_batch` may reject geometry. A false return
+    /// guarantees only that validation accepts every batch, not that prediction
+    /// or duration conversion is infallible. Admission must still protect those
+    /// later operations for external providers.
     fn prefill_batch_validation_can_fail(&self) -> bool {
         true
     }

crates/core/src/perfmodel/config.rs

@@ -105,7 +105,10 @@ pub const ENGINE_CONFIG_SCHEMA_VERSION: u32 = 1;
 // - 19 (DeepSeek-V4.1 review): Dsv41AttentionOp gained kv_cache_layout,
 //   separating physical backend KV payload from attention arithmetic precision.
 //   Its appended enum changes positional bincode layout; old JSON defaults only.
-pub const ENGINE_SPEC_SCHEMA_VERSION: u32 = 19;
+// - 20 (DeepSeek-V4.1 FPM): FpmForwardOp gained original_fmha_quant_mode
+//   for selector diagnostics. This appends a positional field after the schema-19
+//   release; serde defaults support legacy JSON, not legacy bincode.
+pub const ENGINE_SPEC_SCHEMA_VERSION: u32 = 20;
 
 /// Static engine identity and setup information carried by an
 /// [`crate::perfmodel::engine::spec::EngineSpec`].
@@ -254,6 +257,9 @@ pub struct QuantizationConfig {
     #[serde(default)]
     pub moe_dtype: Option<DataType>,
     pub activation_dtype: Option<DataType>,
+    /// FPM cell selector only; does not override model arithmetic or memory.
+    #[serde(default)]
+    pub fpm_fmha_dtype: Option<DataType>,
     pub kv_cache_dtype: Option<DataType>,
 }